TL;DR

• DOJ Joins Federal Challenge to Colorado’s AI Discrimination Law: The U.S. Department of Justice intervened in a federal lawsuit filed by Elon Musk’s xAI against Colorado’s AI discrimination law, marking the first challenge from the federal government to a state AI law in court. The lawsuit argues that the legislation compels developers to embed ideological judgments into AI model outputs and violates the Equal Protection Clause.

• White House Releases Legislative AI Blueprint: The Trump administration issued a national AI policy framework urging Congress to centralize AI oversight and preempt state regulation in order to avoid government censorship. Whether that framing actually safeguards against censorship or merely moves regulatory powers remains an open question.

• Washington State Enacts AI Disclosure and Chatbot Safety Laws: Washington Governor Ferguson signed three AI bills requiring watermarks on AI-generated media, disclosure and safety obligations for companion chatbots, and a new private right of action for unauthorized deepfake uses of a person’s voice or likeness — with critics warning that litigation-driven enforcement may create compliance uncertainty.

• Anthropic Sues Trump Administration Over Pentagon Blacklisting: Anthropic is challenging its designation as a supply-chain risk, alleging the move was direct retaliation over its refusal to allow the Pentagon to use Claude for mass domestic surveillance or autonomous weapons targeting.

• Court Refuses to Block California’s AI Training Data Transparency Law: A federal court declined to halt California’s law requiring AI companies to publicly disclose summaries of their training data, rejecting xAI’s argument that the requirement violates the First Amendment and constitutes an unconstitutional taking of trade secrets.

• Florida Opens Criminal Probe Into ChatGPT Outputs: Florida’s attorney general launched a criminal investigation into OpenAI over allegations that ChatGPT advised the FSU shooting suspect on weapon choice and targeting. The company faces a separate lawsuit from the families of victims of the Tumbler Ridge school shooting in British Columbia, who claim it failed to alert police after internally flagging the shooter’s account months before the attack.

• China Rolls Out National AI Ethics Review Framework: China introduced its first comprehensive AI ethics review regime, requiring pre-deployment approval for AI systems capable of shaping public opinion or social mobilization — placing censorship upstream of any user interaction, with a government-maintained list of high-risk activities that can be updated without notice or challenge.

• India’s AI Content Rules Drive Wave of Takedowns: India’s amended IT Rules have triggered documented removals of journalist and satirist content, while a new draft amendment would require platforms to treat informal ministerial advisories as binding law — raising serious constitutional concerns about overbroad speech regulation.

• Philippines Launches Multi-Agency Deepfake Initiative: The Philippines’ government formalized a whole-of-government response to AI-generated disinformation and deepfakes, combining legal enforcement with platform pressure. Now, legitimate dissent could be criminalized under laws broad enough to cover content “inciting disobedience to lawful authority.”

• Meta Faces Senate Scrutiny Over AI Surveillance Glasses: U.S. senators pressed Meta over plans to add facial recognition to its Ray-Ban glasses, warning the technology could identify people at protests without their knowledge. Civil society groups, including the ACLU and 75 organizations, have also called on Meta to halt the plans entirely, describing them as a threat to anonymous public life.

• African Governments Expand AI-Driven Surveillance: Eleven African governments have spent at least $2 billion on Chinese-built AI surveillance infrastructure to reduce crime, despite researchers finding no evidence that it will do so.— Now, researchers claim it is instead being used to monitor political opponents, journalists, and activists without adequate legal safeguards.

Major Stories

» DOJ Joins Federal Challenge to Colorado’s AI Discrimination Law

• Elon Musk’s xAI sued to block Colorado’s Artificial Intelligence Act, and the Department of Justice moved to intervene—marking the first time the federal government has challenged a state AI law in court.

• Details:

  • xAI’s complaint argues that building an AI model is a form of protected speech under the First Amendment and that compelling developers to redesign systems to avoid disparate demographic outcomes amounts to government-compelled expression.

  • The DOJ, joining through its Civil Rights Division, argues the law also violates the Equal Protection Clause by requiring companies to mitigate disparate impacts while simultaneously carving out an exemption for algorithms designed to advance “diversity” or “redress historical discrimination.”

  • The same afternoon the DOJ intervened, Colorado’s Attorney General agreed to voluntarily halt enforcement of the law pending resolution of the litigation, with the state legislature facing a narrow window to pass an amended version before the current text is either struck down or takes effect on June 30.

Free Expression Implications: The case tests whether AI model design constitutes protected speech and whether anti-discrimination mandates amount to compelled expression. The DOJ argues that a state anti-discrimination law itself distorts expressive outputs. The outcome will shape whether states can mandate how AI systems weigh protected characteristics, and how far First Amendment doctrine extends into AI governance.

For insight into Colorado’s AI Act, see the US chapter of FoFS’s report: “That Violates My Policies: AI Laws, Policies, and the Future of Expression.” The Act regulates “high-risk” AI systems, but exempts chatbots governed by an “accepted use policy” prohibiting “discriminatory or harmful” content, leaving these terms undefined. This could incentivize companies to adopt content restrictions that may sweep in protected speech to avoid liability.

» White House Releases National AI Policy Blueprint

• The White House released its National Policy Framework for Artificial Intelligence, outlining nonbinding legislative recommendations to guide Congress toward a unified federal approach to AI regulation and preemption of state laws.

• Details:

  • The Framework is organized around seven pillars: child protection, AI infrastructure, intellectual property, censorship and free speech, innovation, workforce, and federal preemption of state AI laws.

  • The “Preventing Censorship and Protecting Free Speech” calls on the government to protect free speech and the First Amendment and on Congress to prohibit federal agencies from coercing AI providers to ban or alter content based on partisan or ideological agendas, and to provide redress mechanisms where the government attempts to censor expression on AI platforms.

  • On state preemption, the Framework recommends prohibiting state laws that “impose undue burdens” on AI development.

  • The Framework’s release was met with immediate pushback from the other side of the aisle: Representative Beyer and Democratic colleagues introduced the GUARDRAILS Act, which would nullify the December AI preemption executive order and prohibit the use of federal funds to implement it.

Free Expression Implications: The Policy Framework’s reference to defending the First Amendment is welcome. It is also encouraging that, unlike recent Executive Orders, the framework does not emphasize neutrality or “truth-seeking.” At the same time, previous statements and Executive Orders framed in culture-war terms counsel caution. The Framework’s age-assurance recommendations, which condition access to lawful speech on identity disclosure, also give pause. For additional insight, see our statement here.

» Washington State Enacts AI Disclosure and Chatbot Safety Laws

• Washington Governor Bob Ferguson signed three AI bills in March 2026 covering AI content disclosure, companion chatbot safety, and a new right to sue over unauthorized AI-generated uses of a person’s voice or image.

• Details:

  • House Bill 1170 requires large AI companies to identify when images, video, or audio are created or substantially altered by AI — through watermarks or embedded metadata — and to offer detection tools. House Bill 2225 requires companion chatbots to disclose they are not human at the start of every conversation and every three hours, with hourly reminders for minors, and bans manipulative engagement tactics designed to deepen emotional attachment. A separate law allows individuals to sue over unauthorized deepfake uses of their voice or likeness.

  • Critics warn that allowing enforcement standards to be defined through private lawsuits rather than agency rulemaking may create uncertainty for responsible actors trying to comply in good faith. Ferguson signed the bill without changes.

Free Expression Implications: Disclosure and duty of care requirements can indirectly shape expressive content. Providers facing liability may limit nuanced discussion of sensitive topics — mental health, sexuality, identity — to avoid exposure. The concern is particularly acute for younger users: those who may most need reliable, non-judgmental information on difficult subjects are precisely those for whom the new restrictions apply most stringently.

» Anthropic Sues Trump Administration Over Pentagon Blacklisting, Raising First Amendment Questions

• Anthropic filed a lawsuit in California in March 2026, challenging the Trump administration’s designation of the company as a “supply-chain risk” — effectively barring government agencies from working with Anthropic after the company refused to allow the Pentagon to use Claude for mass domestic surveillance or fully autonomous weapons targeting. Anthropic argues the designation was unconstitutional retaliation for its First Amendment-protected advocacy on AI safety.

• In a parallel proceeding, Anthropic filed in the D.C. Circuit, challenging a separate government-wide supply-chain designation under the Federal Acquisition Supply Chain Security Act, which effectively warned all federal contractors to avoid using Anthropic’s products in defense work.

• Details:

  • In California, a federal judge granted Anthropic a preliminary injunction on March 26, finding the company likely to succeed on its First Amendment arguments and other claims. She called the government’s rationale “Orwellian” and held that the measures looked like punishment for Anthropic’s public criticism rather than a genuine security response. The ruling temporarily blocks the government’s actions, with the government’s appeal now before the Ninth Circuit.

  • At the D.C. Circuit, a three-judge panel denied Anthropic’s emergency stay of the separate FASCSA designation on April 8 — explicitly declining to reach the First Amendment merits, and finding that Anthropic’s harms, while real, were primarily financial rather than constitutional in nature. The ACLU and CDT filed an amicus brief in support of Anthropic, arguing its advocacy for AI guardrails is constitutionally protected.

Free Expression Implications: The case raises significant and unresolved questions, notably whether and how ethical limits implemented by AI companies in their models can be enforced against government customers.

» Court Refuses to Block California’s AI Training Data Transparency Law

• A California federal court declined to halt enforcement of California’s AI training data transparency law in March 2026, denying xAI’s request for a preliminary injunction in X.AI LLC v. Bonta — leaving in place a requirement that generative AI companies publicly post summaries of the datasets used to train their systems.

• Details:

  • California’s AB 2013 requires developers of generative AI systems available in that state to disclose the sources and owners of training datasets, whether they contain personal data or copyrighted material, approximate dataset size, and whether data was purchased or licensed. It does not require disclosure of proprietary model weights or system architecture. OpenAI and Anthropic have already posted compliant disclosures.

  • xAI raised three constitutional challenges: that the law compels disclosure of trade secrets, constitutes an unconstitutional taking of property without compensation, and violates the First Amendment by compelling speech. Judge Bernal denied the injunction, finding xAI had not shown it was likely to succeed on any of these claims — but the ruling addresses only the threshold for emergency relief, not the law’s ultimate constitutionality. The case proceeds on the merits.

Free Expression Implications: The case is still ongoing — the court’s denial of a preliminary injunction leaves the constitutional questions unresolved on the merits. xAI’s First Amendment argument, if ultimately accepted, could mean that decisions about what data to use to shape an AI system’s outputs constitute protected expression that the government cannot compel disclosure of.

Subscribe now

» Florida Launches Criminal Investigation Into ChatGPT Outputs

• Florida Attorney General James Uthmeier announced a criminal investigation into OpenAI, alleging ChatGPT advised the man accused of a mass shooting at Florida State University on what ammunition to use, what time of day to strike, and where on campus to find the most people.

• Details:

  • Uthmeier sent subpoenas to OpenAI requesting its policies on responding when users make threats to harm others, and said at a press conference: “If it was a person on the other end of that screen, we would be charging them with murder.” The criminal investigation follows a civil inquiry announced earlier the same month.

  • OpenAI has disputed the characterization, stating ChatGPT “provided factual responses to questions with information that could be found broadly across public sources” and did not encourage illegal activity, and noted the company proactively shared the suspect’s account information with law enforcement after the shooting.

  • A parallel lawsuit in British Columbia, Canada alleged ChatGPT discussed gun violence with the perpetrator of a mass shooting in February 2026, and that OpenAI’s internal systems flagged the account eight months before the shooting, but the company chose not to alert authorities. OpenAI CEO Sam Altman apologized to the community, but the company is expected to face more than two dozen additional suits in the coming weeks.

Free Expression Implications: The probe raises foundational questions about whether AI outputs are more analogous to protected speech or regulated conduct. No U.S. court has yet held a general-purpose AI provider criminally liable for a user’s violence. If such a precedent were established, providers would likely sharply curtail responses on sensitive topics well beyond what existing law requires, with a significant chilling effect on lawful AI-mediated speech.

» China Introduces National AI Ethics Review Framework

• China has rolled out its first comprehensive AI ethics review framework, jointly issued by ten government departments, requiring universities, research institutions, healthcare providers, and companies to establish ethics compliance systems before deploying AI that could pose risks to human dignity, public order, or health.

• Details:

  • At the core is a pre-deployment approval process: institutions must obtain ethics clearance from internal committees before proceeding, with a mandatory additional government-led expert reassessment for high-risk activities — including algorithms capable of shaping public opinion or social mobilization, and systems that significantly affect behavior, emotions, or health.

  • The high-risk list is updated dynamically by government ministries, and a parallel oversight regime allows authorities to suspend or terminate AI projects if risk conditions change during implementation.

Free Expression Implications: The framework’s most significant provision from a free expression standpoint is its designation of systems capable of “shaping public opinion or social mobilization” as high-risk — subject to government-led expert reassessment before deployment. Pre-deployment review of expressive AI systems places censorship upstream of any user interaction, and the dynamic updating of the high-risk list means the boundaries of permissible AI speech can shift without notice or challenge.

For insight into China’s AI policies, see the China chapter of FoFS’s report: “That Violates My Policies: AI Laws, Policies, and the Future of Expression.”

» India’s AI Content Rules Drive Wave of Takedowns Against Journalists and Satirists

• India’s amended IT Rules, which took effect in February 2026 and introduced a three-hour takedown window for AI-generated content, have been accompanied by a documented surge in government-ordered content removals targeting journalists, satirists, and political commentators — with no public explanation of which ministry ordered the actions or why.

• Details:

  • A March 30 draft amendment would go further, requiring platforms to comply with informal ministerial communications — advisories, clarifications, and directives — as a binding condition of safe harbor protection. Platforms that fail to comply with these communications—which are instruments that carry no force of law and require no parliamentary scrutiny—would face liability for all user content. The same draft would extend a Code of Ethics designed for professional publishers to ordinary social media users discussing current affairs.

  • Critics warn the draft reproduces the logic of a provision the Bombay High Court struck down in 2024 for being unconstitutionally vague and overbroad — but broadens the trigger from a single fact-check unit to an indefinite class of executive communications.

Free Expression Implications: Strict regulatory deadlines create structural incentives toward over-compliance. When platforms face liability for delay or misjudgment, lawful but controversial speech becomes collateral damage. The Indian case illustrates how AI safety regimes can become instruments of speech control when paired with punitive intermediary liability frameworks.

For insight into India’s AI policies, see the India chapter of FoFS’s report: “That Violates My Policies: AI Laws, Policies, and the Future of Expression.”

» Philippines Launches Coordinated Government Initiative Against Deepfakes

• The Philippine government has launched a multi-agency initiative under “Oplan Kontra Fake News,” formalizing coordination among justice, communications, and technology agencies to combat AI-generated deepfakes and online disinformation.

• Details:

  • The Department of Justice, Presidential Communications Office, and Department of Information and Communications Technology signed a memorandum of agreement that combines legal enforcement, platform coordination, and public education. Authorities argued that AI-generated false narratives — particularly on economic issues — could trigger panic, distort markets, and undermine institutions, and noted that content may constitute offenses under the Revised Penal Code and the Cybercrime Prevention Act.

  • In a parallel move, the government wrote to Meta demanding expedited takedown protocols for “high-risk content” and a 24/7 enforcement coordination point, warning that failure to comply could prompt regulatory and legal measures. Meta indicated its platforms are committed to supporting the effort.

Free Expression Implications: The framework’s stated commitment to preserving constitutionally protected speech sits alongside enforcement tools broad enough to reach content that ‘incites disobedience to lawful authority.’ When governments can demand expedited platform takedowns of content deemed harmful to ‘economic stability’ or ‘confidence in institutions,’ there is a real risk that speech is restricted simply because it is inconvenient.

» US Senators Press Meta Over AI Surveillance Glasses

• U.S. senators formally pressed Meta over the civil liberties implications of its AI-enabled smart glasses after internal company documents revealed plans to add facial recognition capable of identifying strangers in real time — a feature Meta internally referred to as “Name Tag.”

Details:

• Senators Markey, Wyden, and Merkley sent Meta CEO Mark Zuckerberg a letter in March 2026 warning that the glasses could be used to identify people at political rallies and demanding answers about data retention, database scope, and whether biometric data would be used to train Meta’s AI models. Meta did not respond by the senators’ April 6 deadline.

• In April 2026, the ACLU and 75 organizations called on Meta to halt and publicly disavow the plans, warning the glasses would allow anyone to identify strangers by name at protests, medical clinics, and other sensitive locations — and link those identities to databases containing information on health, habits, and relationships.

Free Expression Implications: Glasses equipped with facial recognition would eliminate any expectation of anonymity in public life. The ability to be identified — and profiled — during a protest, a clinic visit, or an ordinary social interaction creates a chilling effect that operates before any speech is uttered. As the ACLU coalition warns, this is not merely a privacy concern: the erosion of anonymous public presence is a prerequisite condition for free expression, and its loss would fall hardest on those already most exposed to surveillance and retaliation.

» African Governments Accelerate Deployment of AI-Driven Surveillance Systems

• Governments across Africa are rapidly deploying AI-enabled surveillance technologies—including facial recognition and biometric identification—often with minimal legal oversight, raising alarms about protest monitoring and repression.

• Details:

  • Eleven African countries have spent at least $2 billion on Chinese-built “smart city” surveillance infrastructure, including AI-enabled CCTV and facial recognition systems.

  • The systems are being rolled out without adequate legal frameworks to protect human rights, and researchers warn that they are increasingly being used to monitor political opponents, journalists, and activists rather than to address the security threats that were used to justify their procurement.

Free Expression Implications: When individuals know—or suspect—that their movements can be automatically tracked, participation in protests and political discourse becomes riskier. In environments with limited rule of law, AI surveillance enables repression at scale, eroding anonymity and chilling expression long before overt censorship occurs.

The Future of Free Speech in Action

The Future of Free Speech has been reviewing how model behaviors shape freedom of expression and political conversations, as part of its broader commitment to engage with industry, policymakers, academia, and civil society. Anthropic has highlighted our collaboration in their update on election safeguards. This work builds on projects such as That Violates My Policies: AI Laws, Chatbots, and the Future of Expression.

Senior Research Fellow Jordi Calvet-Bademunt was a speaker at a multistakeholder event in Dublin, attended by industry representatives, civil society, and scholars, that focused on digital rights and risk mitigation. He warned against the dangers posed by vague, risk-based regulations and emphasized the need for enforcement guidance that is firmly committed to protecting freedom of expression.

Jordi will speak at the 2026 Internet Communications Governance Forum, to be held in Taipei on May 12–13, 2026. He will discuss developments in international AI legislation and their implications for free speech. The event is organized by Taiwan’s National Communications Commission and the Taipei Computer Association.

Executive Director Jacob Mchangama will speak at the Copenhagen Democracy Summit on May 12, 2026. He will discuss the “AI and Censorship” dilemma.

Share

Isabelle Anzabi is a research associate at The Future of Free Speech, where she analyzes the intersections between AI policy and freedom of expression.

Jordi Calvet-Bademunt is a Senior Research Fellow at The Future of Free Speech and a Visiting Scholar at Vanderbilt University.

TL;DR

• US Government Bars Anthropic Products From Federal Agencies: President Trump ordered all federal agencies to cease using Anthropic’s AI technology after the company refused to allow unrestricted military use of its models. Defense Secretary Hegseth designated Anthropic a supply-chain risk to national security, barring Pentagon contractors from doing business with the company. The escalation raises serious questions about AI-enabled surveillance and its consequences on speech.

• States Push Broad AI Chatbot Regulation: Lawmakers in more than 35 states are advancing nearly 100 bipartisan proposals to regulate AI chatbots regarding youth mental health, sexual content, deepfakes, and deceptive design. The measures emphasize age verification, parental consent, transparency, and expanded liability with potential impacts on anonymity, privacy, moderation practices, and lawful but sensitive expression.

• White House Pushes Back on State AI Oversight: The Trump administration has signaled broad opposition to state-level AI regulation, most directly by opposing Utah’s HB 286, a bill that would require frontier AI developers to publish safety and child-protection plans. The move reflects growing federal-state tension over who controls the emerging AI governance landscape.

• State AG Scrutiny of AI Chatbots Intensifies: The New Mexico Attorney General’s Office has launched an investigation into how AI chatbots interact with teenagers, as the Federal Trade Commission separately requests information from major AI developers about risks associated with AI companion chatbots. The combined activity signals escalating state and federal scrutiny of AI companions and growing tension over youth safety rules.

• India Tightens AI Takedown Rules’ Timeline: India’s amended IT Rules drastically shorten the compliance timeline for removing violative AI-generated content, expand labeling obligations, and broaden the prohibited categories. The changes heighten the risk of precautionary takedowns, increasing pressure on lawful speech, including satire and political commentary.

• Grok Draws Coordinated Global Action Over Sexual Deepfakes: xAI’s Grok chatbot is under investigation across Brazil, the UK, the EU, India, and South Korea over non-consensual and sexualized AI-generated imagery. Regulators are converging on an approach that treats generative AI tools embedded in platforms like X as accountable intermediaries under existing digital safety, data protection, and child protection regimes.

• EU Prepares for AI Act Enforcement in 2026: The European Commission is pivoting from rulemaking to enforcement, prioritizing oversight of general-purpose AI models, regulatory sandboxes, and AI Office market surveillance. Deferred definitions, such as systemic risk thresholds, are likely to shape cautious moderation practices for expressive AI outputs.

• ICE’s Use of Facial Recognition Extends Into Protest Settings: Reporting shows U.S. Immigration and Customs Enforcement and DHS deploying mobile facial recognition tools, including Mobile Fortify, during operations that intersect with public demonstrations, prompting legal challenges and renewed concerns about chilled speech, anonymity, and association.

• Iran Scales AI-Driven Surveillance to Suppress Dissent: Iran has expanded AI-enabled monitoring and facial recognition—reportedly with support from Chinese technology providers—to identify protesters and track online activity, illustrating how integrated biometric systems can institutionalize repression and erode expressive freedom.

• Meta Limits AI Reproductive Health Information for Teens: Leaked internal policies show that Meta’s AI chatbots restrict users under 18 from receiving information about abortion and other sexual and reproductive health topics, including neutral educational content, with the limits applied globally, even in jurisdictions where such information is not under legal threat.

• OpenAI Partners with the UAE on Censored ChatGPT Version: OpenAI is building a localized ChatGPT for the United Arab Emirates fine-tuned for the local Arabic dialect and calibrated to reflect the monarchy’s political line.

Major Stories

» US Bars Anthropic Products From Federal Agencies and Contractors

• President Trump ordered all federal agencies to cease using Anthropic’s AI technology after the company refused to grant the Pentagon unrestricted use of its models.

• Defense Secretary Pete Hegseth designated Anthropic a supply-chain risk to national security, barring Pentagon contractors and their partners from any commercial activity with the company.

• Details:

  • Background: Anthropic insisted that Claude not be used for mass surveillance against Americans or in fully autonomous weapons operations. The Pentagon demanded unrestricted access for all lawful purposes, setting a compliance deadline that expired without agreement.

  • Supply-Chain Designation: The supply-chain risk label, typically reserved for adversarial foreign entities, bars any military contractor or supplier from doing business with Anthropic. The designation also affects other firms that deploy Anthropic’s technology in their own products, including Palantir’s Maven Smart System battle management platform.

  • Industry Ripple Effects: OpenAI quickly struck a deal with the Pentagon to deploy its models on classified networks. Elon Musk’s xAI recently received approval for classified work as well. Employees at several major tech companies, including Amazon and Microsoft, called for their firms to reject Pentagon demands for unrestricted AI use. On Saturday, Claude overtook ChatGPT to become the number one free app in Apple’s App Store.

  • Legal Challenge: Anthropic has stated it will challenge the supply-chain risk designation in court, arguing the move is legally unsound and sets a dangerous precedent for any American company that negotiates with the government.

  • Continued Military Use: Despite Trump’s order to sever ties with Anthropic, the US military reportedly continued using Claude during the joint US-Israel bombardment of Iran that began on Saturday. According to the Wall Street Journal and Axios, military command used Claude for intelligence purposes, target selection, and battlefield simulations. This follows the earlier use of Claude in the US raid to capture Venezuelan President Nicolás Maduro in January. Defense Secretary Hegseth acknowledged the difficulty of rapidly detaching military systems from the AI tool, granting a six-month transition window.

Free Speech Implications: From a free speech perspective, concerns about AI-enabled surveillance warrant serious consideration. Governments around the world are already leveraging AI to monitor their populations more effectively, and the expansion of such capabilities raises legitimate worries about chilling speech. When people know they may be watched — or suspect they are — they self-censor, and that dynamic strikes at the heart of free expression. This is unfolding in the context of the current administration’s war against “woke AI,” which risks substituting one orthodoxy for another.

» States Advance Sweeping AI Chatbot Regulation Push

• Lawmakers in at least 35 states are considering nearly 100 bills that would regulate AI chatbots, citing concerns about youth mental health, deepfakes, and transparency. The proposals span both Republican and Democratic legislatures, with safety and age restrictions emerging as bipartisan priorities.

• Details:

  • Safety & Crisis Protocols: More than 25 states are weighing bills requiring chatbot platforms to implement suicide-prevention and crisis response systems. Pennsylvania and Louisiana are considering measures to restrict or ban chatbots from serving as substitutes for licensed therapists, following similar laws passed in Tennessee, Illinois, and Nevada.

  • Deepfake & Sexual Content Safeguards: State investigations into Elon Musk’s Grok chatbot over alleged nonconsensual sexualized deepfakes have intensified scrutiny. Bills in California and Arizona would strengthen protections against exposing minors to sexually explicit AI-generated material.

  • Transparency Requirements: At least 20 states are advancing disclosure mandates requiring platforms to inform users when they are interacting with a chatbot. Washington’s SB 5984 has already cleared the state Senate.

  • Age Verification & Parental Controls: More than 20 proposals would require chatbot platforms to verify users’ ages. Some states would require parental consent for minors, while others would restrict “human-like” chatbot features for underage users.

  • Expanded Liability: Several bills create private rights of action or civil penalties. For instance, Illinois lawmakers have proposed classifying chatbots as products under strict liability law, thereby increasing legal exposure if users are harmed.

Free Speech Implications: The proposals shift regulatory focus from user speech to AI system design and deployment. Age-verification mandates, if upheld by courts, will lead to the erosion of anonymity online. Expanded liability and crisis-monitoring requirements could incentivize platforms to limit conversational capabilities or over-moderate lawful but sensitive discussions.

» Utah Transparency Bill Draws White House Scrutiny

• The Trump administration has signaled broad opposition to state-level AI oversight efforts, most starkly illustrated by a White House letter to Utah lawmakers declaring its opposition to HB 286, a bill that would require frontier AI developers to publish safety and child protection plans.

• Details:

  • Utah lawmakers have advanced legislation requiring AI systems to disclose when users are interacting with artificial intelligence rather than a human.

  • Reporting indicates the White House stated its opposition regarding the measure to Utah officials, reflecting broader federal-state tensions over AI regulation.

  • Separately, the Federal Trade Commission has used its Section 6(b) authority to request information from major AI developers about how they assess and mitigate risks associated with AI companion chatbots.

Free Speech Implications: The White House’s direct intervention against a state AI safety bill raises questions about preemption and centralized federal control over emerging AI governance, potentially limiting states’ ability to set independent guardrails on AI systems.

» New Mexico AG Investigates AI Chatbots

• New Mexico Attorney General Raúl Torrez has launched an investigation into AI chatbot platforms to examine how they interact with teenagers and young users, citing concerns about product design and potential harm.

• Details:

  • The New Mexico Department of Justice is seeking information from AI chatbot companies about how their systems engage minors and what safety mechanisms are in place.

  • Torrez said the inquiry marks the beginning of a broader effort to hold technology companies accountable.

Free Speech Implications: Investigations into chatbot design shift regulation toward the structure of AI-mediated communication. Requirements tied to youth safety may influence how conversational AI systems operate, potentially incentivizing platforms to limit or over-moderate lawful but sensitive discussions with younger users.

» India Reduces Time to Comply With AI Takedown Orders to Three Hours

• The Indian government’s amendments to its IT Rules impose a three-hour deadline for platforms to act on official takedown orders related to violative AI-generated media, and a two-hour deadline after notification of non-consensual intimate imagery. Platforms risk losing intermediary liability protections if they fail to comply.

• The revisions also mandate:

  • Disclosure requirements for synthetic content and robust verification mechanisms.

  • Labeling and traceability features for deepfakes and other synthetic media.

  • Expansion of prohibited content categories to include deceptive impersonation, non-consensual intimate imagery, and serious-crime-linked material.

  • Mechanisms enabling platforms to disclose user identities to private complainants in some instances.

Free Speech Implications: The compressed response windows create immense operational pressure on intermediaries, encouraging risk-averse content takedowns. Expanded prohibited categories intersect with expressive norms around satire, political critique, and identity expression, increasing risks that lawful speech will be included in enforcement actions.

» “Grok” Faces Expanding Global Scrutiny Over Sexual Deepfakes

• xAI’s Grok chatbot, integrated into X, is under investigation in multiple jurisdictions following reports that it generated sexualized synthetic images, including non-consensual depictions of real individuals and minors.

• Details:

  • Brazil: Brazilian authorities issued a joint recommendation requiring X to implement, within 30 days, procedures to identify, review, and remove sexualized or eroticized AI-generated content, particularly involving children or non-consenting adults.

  • United Kingdom: The UK Information Commissioner’s Office opened an inquiry into X over allegations that Grok generated non-consensual sexual imagery. The move comes as the UK accelerates efforts to strengthen criminal penalties for sexually explicit deepfakes.

  • European Union: European Commission officials have indicated that X could face scrutiny under the Digital Services Act if it fails to meet risk mitigation and transparency obligations related to AI-generated content.

  • India & South Korea: Indian authorities are reviewing the platform. South Korean regulators have also signaled they are examining explicit AI outputs under domestic digital safety laws.

  • Also in Europe: Spain’s cabinet has approved draft legislation strengthening protections against AI-generated deepfakes, including setting 16 as the minimum age to consent to the use of one’s image and updating penalties for non-consensual synthetic media.

Free Speech Implications: The Grok investigations illustrate how regulators are increasingly treating generative AI systems as accountable speech intermediaries. By layering child protection, data protection, and platform governance rules onto AI outputs, governments may push providers toward stricter, uniform content controls.

» EU Commission Sets 2026 AI Act Implementation Priorities

• Brussels is transitioning from lawmaking to enforcement under the Artificial Intelligence Act, with the EU Commission outlining a 2026 implementation roadmap that prioritizes oversight of general-purpose AI models, regulatory sandboxes, and coordinated market surveillance while deferring some unresolved definitional thresholds.

• Details:

  • General-Purpose AI Oversight: Monitoring compliance for foundational AI models that power chatbots and generative tools, including verifying documentation, transparency standards, and safety reporting.

  • Regulatory Sandboxes: Rolling out controlled environments where innovators can test AI systems under supervision.

  • AI Office & National Coordination: Strengthening the AI Office’s market surveillance role and aligning national supervisors to avoid fragmented enforcement.

  • At the same time, some definitional criteria, such as “systemic risk” thresholds and energy-use methodologies, are reportedly being deprioritized in 2026 internal documents.

• Additionally, the EU’s Anti-Racism Strategy 2026-2030 commits the Commission to providing guidelines on AI Act prohibitions targeting racially biased AI practices, including manipulative AI systems that incentivize hatred or violence against protected groups.

Free Speech Implications: As enforcement begins, procedural definitions will shape how AI systems handle expressive outputs that touch on politics, culture, or identity. Operational guidance, including the General-Purpose AI Code of Practice, could influence moderation defaults, potentially pushing providers toward cautious removal of ambiguous content to avoid regulatory exposure.

» ICE’s Facial Recognition Deployment Raises Protest Surveillance Concerns

• The use of facial recognition technology by the U.S. Immigration and Customs Enforcement and Department of Homeland Security agents has been documented and challenged in Maine.

• Details:

  • This purportedly includes Mobile Fortify, which allows agents to scan faces in real time and match them against biometric databases containing hundreds of millions of images.

  • This capability has been deployed alongside Operation Metro Surge, a federally led immigration enforcement campaign that has included raids and public protests.

  • The use of biometric scanning has prompted legal challenges from state attorneys general and civil liberties advocates, who argue it expands surveillance beyond traditional immigration enforcement contexts into protest environments.

Free Speech Implications: As we’ve previously covered, mobile biometric surveillance during public demonstrations can chill participation and undermine anonymity. When law enforcement uses AI-enabled identification tools without clear judicial or legislative limits, individuals may self-censor out of fear of being catalogued, threatening both speech and association rights.

» Iran’s Use of AI-Enabled Facial Recognition to Track Protesters

• Iran is increasingly integrating AI-assisted surveillance systems, including facial recognition technology, into its broader censorship and public security architecture, enabling authorities to identify individuals involved in protests and control digital communications.

Details:

• Imported Chinese Surveillance Tools: According to recent reports, Iran’s Internet control infrastructure incorporates technology supplied under agreements with Chinese firms, including facial recognition tools repurposed for Iranian security operations. These systems were part of the state’s censorship apparatus, which enabled near-total internet shutdowns during nationwide anti-government protests in early 2026.

• Biometric Identification Capacity: The imported surveillance hardware reportedly allows security forces to identify individuals via cameras on the ground and match them against state databases, suggesting the potential for real-time or retrospective identification of protest participants.

Free Speech Implications: AI-enabled biometric identification significantly reduces the anonymity traditionally associated with public protest and political dissent. When facial recognition systems are integrated with national databases and real-time analytics, participation in demonstrations or online activism could become immediately traceable. Cross-border transfers of AI surveillance technology may normalize automated repression tools, accelerating the global diffusion of infrastructure capable of chilling political expression at scale.

» Meta’s AI Chatbot Limits Reproductive Health Information for Teens

• Leaked internal policy documents show that Meta’s AI chatbots have been configured to restrict what users under 18 can be told about abortion and reproductive health topics.

• The policies emerge amid intensified political pressure on abortion access across many U.S. states, while Meta has previously faced criticism for being too permissive toward minors on sexual content across its platforms.

• Details:

  • According to internal guidelines obtained by Mother Jones, the restrictions bar chatbots from providing teens with information, advice, or opinions on how to locate abortion services, access or use contraception, understand puberty or menstrual health, prevent sexually transmitted infections, or discuss sexual consent.

  • These limits extend beyond actionable guidance to include neutral informational content.

  • These restrictions also apply in jurisdictions where abortion access and information are legal and not under regulatory threat.

Free Speech Implications: By excluding basic reproductive and sexual health information from AI responses to teens, Meta’s internal rules can limit access to lawful, educational content at a stage when many young people seek health information online. This raises concerns about how corporate content policies shape what topics automated systems will allow users to explore, especially on health matters that are legally protected and socially significant.

» OpenAI Developing Censored ChatGPT for UAE Government

• OpenAI is working with Abu Dhabi-based AI conglomerate G42 to build a fine-tuned version of ChatGPT for the United Arab Emirates, designed to accommodate local Arabic dialect, political outlook, and speech restrictions, according to Semafor. The customized chatbot is being designed for use by the UAE government and would represent one of the first regional implementations of a leading AI chatbot.

• Details:

  • Localized Fine-Tuning: The project involves post-training adjustments to an existing ChatGPT model to make it fluent in the local Arabic dialect. Content restrictions are expected to align the chatbot’s outputs with Emirati law and societal values, with sources indicating the UAE intends the chatbot to reflect a political line consistent with the monarchy’s positions. This approach could particularly affect vulnerable communities in the country, including LGBTQ individuals.

  • G42’s Role and Ties: G42 is chaired by Sheikh Tahnoon bin Zayed Al Nahyan, an Abu Dhabi royal who also serves as national security adviser and leads the country’s largest sovereign wealth fund. G42’s investment arm, MGX, invested in OpenAI in a transaction valuing the company at $500 billion, and G42 is also part of the Trump administration’s Stargate AI infrastructure project.

  • Global Version Remains Available: OpenAI plans to continue offering the standard global ChatGPT in the UAE, with adjustments to comply with Emirati law when certain content is prohibited. In those cases, users would be informed when their requests trigger legal restrictions.

Free Speech Implications: The creation of a government-aligned, regionally restricted version of a leading AI chatbot illustrates how localization can become a mechanism for embedding state censorship directly into conversational AI infrastructure. By fine-tuning a chatbot to reflect a monarchy’s political line and local speech restrictions, the project risks normalizing the practice of building government-approved information filters into AI systems. As similar localization models are adopted in other jurisdictions, they could set precedents for how governments shape the expressive boundaries of AI-mediated communication, limiting users’ access to information and perspectives that fall outside state-sanctioned narratives.

The Future of Free Speech in Action

Senior Research Fellow Jordi Calvet-Bademunt delivered a session at the India AI Impact Summit, where he presented the key findings of our recent report on AI laws, chatbots, and freedom of expression and its implications for discussions of global AI governance.

Senior Legal Fellow Ashkhen Kazaryan presented on “Free Speech in the Age of AI: Who Controls the Future of Expression?” and did a Q&A session at Southwestern Law School.

Isabelle Anzabi is a research associate at The Future of Free Speech, where she analyzes the intersections between AI policy and freedom of expression.

Jordi Calvet-Bademunt is a Senior Research Fellow at The Future of Free Speech and a Visiting Scholar at Vanderbilt University.

Share

TL;DR

• President Trump Signs “One-Rule” AI Order Targeting State Laws: The Trump administration has issued an executive order directing federal agencies to challenge state AI regulations and advance a “minimally intrusive national standard,” empowering the Attorney General, FTC, and other agencies to move against laws that require AI models to alter “truthful outputs” or impose disclosure mandates seen as violating the First Amendment — potentially preempting state rules on political deepfakes, algorithmic transparency, or other speech-related regulation.

• European Parliament Pushes for EU-Wide Age Limits on AI and Social Media: The European Parliament backed a non-binding call for a uniform under-16 restriction on social media and AI chatbots, raising concerns about restricting minors’ access to information and online expression.

• EU Proposes Digital Omnibus Regulation: The European Commission unveiled a draft Digital Omnibus Regulation updating transparency, reporting, and content moderation rules for online platforms. The law aims to reduce compliance burdens and harmonize requirements across the EU, but could still incentivize over-moderation and influence what users feel safe posting.

• South Korea Releases Draft Enforcement Decree for AI Basic Act: The Ministry of Science and ICT unveiled rules for high-performance AI, generative content labeling, and industry support, opening a public consultation ahead of the Act’s January 2026 implementation.

• Turkey Proposes AI Legal Framework With Liability and Takedown Rules: A draft law would define “AI system,” impose developer and user liability, mandate labeling and rapid takedowns of harmful content, and grant authorities broad powers to block AI outputs — raising risks of over-censorship and chilling lawful expression.

• Brazil Proposes National AI Governance Framework: President Lula has sent a bill to Congress that would create a coordinated AI governance system led by the data protection authority, with binding transparency standards and sectoral oversight — raising concerns about over-compliance that could indirectly chill journalistic, artistic, and political uses of AI.

• Rights Groups Demand Limits on Facial‑Recognition Use: Civil‑liberties organizations call on DHS to suspend ICE’s Mobile Fortify app and highlight NYPD’s biased use of facial recognition, warning that current practices risk misidentification, target marginalized communities, and could chill lawful expression.

• TikTok Adds Controls to Limit AI Content Amid Surge of Unlabeled Deepfakes: TikTok has introduced new settings to reduce or opt out of AI-generated videos after research found billions of views for unlabeled synthetic content, including misleading political and sexualized posts.

Major stories

» President Trump Signed a “One Rule” AI Executive Order Challenging States’ AI Laws

• President Donald Trump has signed an executive order directing federal agencies to challenge state AI regulations and push toward a minimally burdensome national standard, after lawmakers dropped efforts to embed preemption language in a must-pass defense bill.

• Details:

  • President Trump signed Ensuring a National Policy Framework for Artificial Intelligence, an executive order intended to prevent a patchwork of 50 different state AI regulatory regimes from impeding AI innovation.

  • The order directs the Attorney General to create an AI Litigation Task Force within 30 days to challenge state laws viewed as conflicting with federal policy, including under constitutional and preemption arguments.

  • Within 90 days, the Secretary of Commerce must evaluate state AI laws and identify those that “require AI models to alter their truthful outputs” or compel disclosures that would violate the First Amendment or other constitutional rights.

  • The FTC must issue a policy statement clarifying when state laws that require changes to truthful AI outputs are preempted by the FTC Act’s bar on deceptive practices.

  • The FCC is instructed to consider a federal reporting and disclosure standard for AI models that would supersede conflicting state requirements.

  • The Secretary of Commerce must also outline when states may remain eligible for Broadband Equity, Access, and Deployment (BEAD) funding, withholding certain funds from states with “onerous AI laws,” and agencies must review grant programs for potential similar conditions.

• Background:

  • Congressional efforts to secure preemption via the must-pass defense bill were recently formally abandoned: House Majority Leader Steve Scalise stated that the language would not be included. “If you can add [preemption language], that’s great, but that wasn’t the best place for this to fit.”

Free Speech Implications:

• Several states have pursued AI regulations at odds with free speech principles, such as laws banning political deepfake passed in California and Minnesota.

• However, the emphasis on “truthful outputs,” rather than on core First Amendment grounds, raises questions about whether this effort is aimed more at advancing specific viewpoints favored by the government than at providing a principled protection of free speech.

» European Parliament Pushes for EU-Wide Minimum Age to Access AI Chatbots & Social Media

• The European Parliament approved a non-binding resolution recommending that the Digital Services Act (DSA) framework be supplemented with a uniform EU-wide minimum age of 16 for access to social-media platforms, video-sharing sites, and AI chatbots/companions.

• Details:

  • The proposal would allow individuals aged 13–16 to access those services only with parental consent, and would entirely bar under-13s from access.

  • Parliament calls for restrictions on addictive design features for minors, the addressing of AI tools capable of generating fake or inappropriate content, and banning non-compliant websites.

  • The resolution has no legal force and expresses Parliament’s position on the matter.

Free Speech Implications: This resolution aligns with the Australian government’s ban of under-16s from accessing social media. The resolution could carry negative impacts on minors’ access to information and expression online.

» EU Unveils “Digital Omnibus” Proposal to Streamline AI and Digital Rules

• The European Commission released its Digital Omnibus proposal, a wide-ranging update to EU digital laws meant to reduce compliance burdens and align deadlines across data, AI, and cybersecurity frameworks.

• Details:

  • The package is part of a broader simplification agenda aiming to make digital rules “clearer and more coherent” for businesses.

  • For AI, the Omnibus would delay several EU AI Act deadlines, including postponing requirements for high-risk AI systems until harmonized standards are in place and pushing the transparency obligation for AI-generated content to February 2027.

  • It would also streamline conformity assessments, expand flexibilities for SMEs and mid-caps, and centralize elements of AI oversight within the EU AI Office to reduce duplication across member states.

  • The Commission argues the changes will help companies comply more effectively, but rights groups warn the package could dilute safeguards embedded in existing data-protection and AI-governance rules.

Free Speech Implications:

• Clearer requirements on transparency and reporting could help platforms justify decisions to users, supporting accountability and potentially reducing inconsistent content removal.

• While bans on political deepfakes raise significant free speech concerns, the discussion surrounding disclosure—its implications for free speech, effectiveness, and feasibility—remains ongoing.

» South Korea Releases Draft Enforcement Decree for AI Basic Act

• South Korea’s Ministry of Science and ICT released the draft enforcement decree for the AI Basic Act, its first comprehensive AI legislation, and opened a public consultation period. The decree clarifies standards and obligations under the Act.

• Key Measures:

  • Establishes standards for support projects aimed at fostering the AI industry.

  • Designates institutions responsible for implementing national AI policy.

  • Defines “high-performance AI model” based on computational power, and such models are subject to additional safety regulations and oversight.

  • Introduces disclosure and labeling requirements for “deepfakes” and other generative content to ensure users are properly informed.

  • Confirms exemptions for AI used exclusively for national defense or security purposes.

  • The decree reflects input from the National AI Strategy Committee, industry stakeholders, and experts, and is expected to be finalized before the AI Basic Act takes effect in January 2026.

Free Speech Implications: Disclosure and labeling requirements for generative content may help users understand AI outputs, but compliance obligations could encourage cautious or conservative deployment, potentially affecting expressive, artistic, or political uses of AI.

» Turkey Proposes Broad AI Legal Framework With Liability, Takedown, and Labeling Rules

• A draft law submitted to Turkey’s Grand National Assembly would introduce a legal definition of “AI system,” clarify liability for users and developers, and impose new obligations and enforcement mechanisms across multiple statutes, marking a significant expansion of state control over AI‑generated content and technology.

• Details:

  • The bill would amend Turkey’s Internet Law (Law No. 5651), Turkish Penal Code No. 5237, and other laws to define “AI system” as software or models that autonomously or semi‑autonomously process data to generate outputs.

  • It creates liability for individuals who direct an AI to commit conduct that constitutes a crime, treating them as perpetrators, and increases penalties for developers whose design or training of AI enables such acts.

  • Content generated by AI that violates personal rights, endangers public security, or involves deepfakes must be blocked or removed within six hours, with both content providers and AI developers jointly liable for compliance.

  • AI‑generated material must carry a visible label with administrative fines for failure to disclose and potential access blocking orders for systematic non‑compliance.

  • The draft would empower the Information and Communication Technologies Authority (BTK) to impose fines and issue urgent blocking orders for AI content that threatens public order or election security.

Free Speech Implications:

• The strict takedown timeline and joint liability expose platforms and developers to risk, likely encouraging over‑removal of user content to avoid penalties — including lawful political or cultural speech.

• Mandatory labelling and rapid enforcement may stifle creative, journalistic, and satirical uses of AI, particularly where expressive nuance or context matters.

• Broad emergency blocking power over “threats to public order or election security” — undefined in the bill — could enable arbitrary enforcement against dissenting voices, chilling public debate and minority expression.

• In July, a Turkish court partially blocked access to xAI’s chatbot Grok, which the court found responsible for generating content “insulting” to both President Erdogan and modern Turkey’s founder, Kemal Ataturk, as well as religious values.

Subscribe now

» Brazil Sends Sweeping AI Governance Bill to Congress

• President Luiz Inácio Lula da Silva has submitted a comprehensive AI governance bill to Brazil’s Congress. Bill No. 6237/2025 would create the National System for the Promotion and Regulation of Artificial Intelligence, bringing together the Brazilian Council for Artificial Intelligence (CBIA), the National Data Protection Authority (ANPD), sectoral regulators, and advisory bodies involving civil society and experts.

• Details:

  • The bill establishes a coordination framework for AI policy and oversight across government, rather than setting detailed content or speech rules.

  • CBIA would define national AI policy and strategic priorities, while ANPD would act as the central coordinator, issuing binding transparency standards and general rules.

  • Sectoral regulators would oversee AI uses within their domains under these cross-cutting norms.

  • While the proposal references risk-based oversight, specific high-risk classifications and obligations remain part of Brazil’s broader AI legislative debate.

Free Speech Implications:

• Vague transparency requirements and sectoral oversight of high-risk systems may encourage over-compliance, affecting journalistic, artistic, and political uses of AI.

• To learn more about Brazil’s AI regulations and freedom of expression, read Carlos Affonso Souza’s chapter from our report — That Violates My Policies: AI Laws, Chatbots, and the Future of Expression.

» Civil Liberties Groups Demand End to Facial-Recognition Use by ICE & NYPD

• Civil-liberties organizations are calling for an immediate halt to two separate U.S. law-enforcement agencies’ use of mobile facial-recognition technology (FRT)— citing documented abuses that threaten privacy, protest rights, and free expression.

• ICE’s Mobile Fortify App — National Biometric Surveillance on the Street

  • Civil-liberties organizations, including the Electronic Frontier Foundation (EFF), Asian Americans Advancing Justice and the Project on Government Oversight, are urging the Department of Homeland Security (DHS) to immediately stop Immigration and Customs Enforcement (ICE)’s use of “Mobile Fortify,” a handheld FRT that allows agents to take a photo during an encounter and run it across DHS databases, pulling immigration files and other identity data.

  • The coalition warns the tool enables warrantless, suspicionless biometric scanning of immigrants, U.S. citizens, and bystanders, with no public oversight.

  • They also cite DHS’s history of biometric misidentification as evidence that the tool risks wrongful stops or detentions, with data potentially retained long-term and shared broadly.

• NYPD’s Documented Use of FRT Against Protesters and Communities of Color

  • Newly released New York Police Department (NYPD) records obtained by Amnesty International and the Surveillance Technology Oversight Project show extensive and discriminatory use of FRT, including repeated targeting of protesters, people of color, and individuals engaged in constitutionally protected activity.

  • The records show that NYPD analysts routinely used FRT in ways that risked racially biased identification, including flagging people based on facial features and hairstyles stereotypically associated with communities of color.

  • Images used for matching often came from social media, protest footage, or low-quality surveillance feeds, increasing the likelihood of misidentification.

Free Speech Implications:

• EFF warns that pervasive, on-the-spot biometric scanning could chill participation in protests, public gatherings, and other expressive activities, especially in communities already subject to heightened immigration enforcement.

• Because both ICE and NYPD deploy facial recognition in public and street-level contexts — often without consent, transparency, or oversight — the technologies pose a serious threat to anonymity, protest, and free expression. For marginalized communities, the racial and ethnic biases documented in facial-recognition tools add a further layer of risk.

» TikTok Addresses Surge of AI-Generated Content With New User Controls

• TikTok rolled out new settings that let users reduce or opt out of AI-generated content in their feeds, as new research shows the platform is being flooded with unlabeled synthetic media, including anti-immigrant narratives and sexualized content, which has reached billions of views.

• Details:

  • TikTok announced that users can now “tune down” or opt out of AI-generated videos in their For You Page and will see clearer “AI-made” labels on realistic synthetic media.

  • The update comes amid mounting pressure after AI Forensics, a Paris-based non-profit, found hundreds of AI-driven accounts pushing nearly 43,000 AI-generated posts per month, much of it unlabeled.

  • The report identified content with anti-immigrant messaging, sexualized depictions, and misleading “news-style” posts — collectively garnering billions of views.

  • TikTok has labeling policies but, in practice, enforcement remains inconsistent and many AI-generated videos remain undetected.

Free Speech Implications: New user controls will give individuals more autonomy over their feeds, but they also raise concerns that platforms could algorithmically shadow-ban AI-generated or AI-assisted artistic or political content.

Share

The Future of Free Speech in Action

The Future of Free Speech convened an international roundtable in Sweden, in collaboration with Örebro University, to examine how artificial intelligence is reshaping freedom of expression. Building on our flagship report That Violates My Policies: AI Laws, Chatbots and the Future of Expression, the event brought together leading academics, regulators, companies, and civil-society experts to explore how generative AI is transforming creativity, influencing democratic dialogue, and redefining the boundaries of lawful speech.

Senior Research Fellow Jordi Calvet-Bademunt delivered a session on AI and freedom of expression to employees of Orange in celebration of Human Rights Day. Orange is one of France’s largest telecommunications operators, and the session explored the implications of recent public policies globally and corporate practices in the AI space.

Jordi will also discuss the EU AI Act and the Digital Services Act, and their implications for freedom of expression, in conference GenAI & Creative Practices: Past, Present, and Future at the University of Amsterdam on December 18.

Executive Director Jacob Mchangama participated in a Lawfare podcast —Scaling Laws: AI Chatbots and the Future of Free Expression with Jacob Mchangama and Jacob Shapiro — discussing the findings of our report and AI and free speech broadly.

Isabelle Anzabi is a research associate at The Future of Free Speech, where she analyzes the intersections between AI policy and freedom of expression.

Hirad Mirami is a research assistant at The Future of Free Speech and a student at the University of Chicago studying economics and history.

Jordi Calvet-Bademunt is a Senior Research Fellow at The Future of Free Speech and a Visiting Scholar at Vanderbilt University.

TL;DR

• New Report Tests Chatbots and Laws on Free Speech: The Future of Free Speech’s report ranks major AI models and laws by their handling of contentious topics. Grok 4 and GPT-5 were most open; China’s DeepSeek and Qwen were most restrictive. The U.S. ranked first for protecting AI-related free speech, with the EU a close second, while China ranked last due to a web of state-imposed content controls.

• Dutch Probe Finds Chatbots Gave Biased Election Advice: The Dutch privacy regulator reported that OpenAI’s GPT-5, xAI’s Grok 4, and Mistral 3.1 provided skewed voting recommendations ahead of national elections, potentially breaching the EU AI Act — risking government overreach regarding political advice.

• California and Congress Target AI Companions for Minors: California became the first state to regulate AI chatbots designed for emotional companionship, requiring age verification, transparency, and mental health safeguards. In Congress, a bipartisan bill proposes a nationwide ban on AI “companions” for minors, which would limit minors’ expressive rights.

• UN Declaration on AI and Expression: UN rapporteurs issued a joint declaration urging both protection of free expression across the AI lifecycle and stronger measures against hate speech/disinformation — a tension that could fuel overbroad moderation.

• India Floats AI Labelling Rules: India’s MeitY proposed IT Rule amendments would require visible labels and metadata for synthetic content — a transparency move that risks stigmatizing and chilling legitimate expressive uses unless narrowly tailored.

• Robby Starbuck Sues Google: Conservative activist Robby Starbuck sued Google over AI products that allegedly generated defamatory claims. The case (after a prior settled suit versus Meta) could reshape liability for AI-generated falsehoods and influence platform moderation behaviors.

• Sora 2 Deepfake Controversy: OpenAI’s Sora 2 generated videos of deceased public figures, prompting a pause on some depictions and sparking debate over platform policy and the balance between creative free expression and publicity rights.

Major Stories

» New Report Tests Chatbots and AI Laws on ‘Free Speech Culture’

• A new study by The Future of Free Speech examines how public and private systems of governance shape the ways generative AI influences free expression and access to information worldwide.

• The report, That Violates My Policies: AI Laws, Chatbots, and the Future of Expression, ranks six major jurisdictions by how they protect free expression in AI legislation and eight leading chatbots by how they handle contentious topics.

• Findings:

  • xAI’s Grok 4 topped the ranking for openness to contested speech, followed by OpenAI’s GPT-5, Anthropic’s Claude Sonnet 4, Google’s Gemini 2.5 Flash, Meta’s Llama 4, and Mistral Medium 3.1.

  • China’s DeepSeek and Alibaba’s Qwen were found to be the most restrictive, routinely avoiding politically sensitive topics.

  • Across all models, “hard moderation” (outright refusals to generate content) is down, but “soft moderation” — steering users away from difficult subjects — remains.

  • The U.S. currently leads the legal ranking for free speech protection in AI, although emerging state laws on “deepfakes” could narrow that lead, and the federal administration’s war on “Woke AI” may further restrict it.

  • While the EU also performed strongly, vague provisions in the AI Act and Digital Services Act risk creating “a culture of self-censorship,” while China’s AI laws hard-code censorship into model design.

Free Speech Implications: “In our tests, some models are starting to engage instead of evade, but the rules shaping those choices are still vague and shifting. If democracies want a pluralist AI ecosystem, both lawmakers and companies need clearer, rights-based guardrails.” — Jordi Calvet-Bademunt

» Dutch Probe Finds Chatbots Gave Biased Election Advice

• The Dutch Data Protection Authority (AP) found that several leading AI chatbots — including OpenAI’s GPT-5, xAI’s Grok 4, and Mistral Medium 3.1 — provided distorted or incomplete voting recommendations ahead of the Netherlands’ parliamentary elections.

• Details:

  • The AP reported that the chatbots overrepresented right-wing PVV and Green-Labor parties, while omitting others entirely. The regulator warned that such skewed “voting advice” could undermine the integrity of democratic processes.

  • Although the AI Act’s enforcement regime won’t take effect until August 2026, the Dutch authority argued that chatbots giving voting advice could be classified as a high-risk system. The findings were shared with the European Commission, and the companies could face early scrutiny or private litigation.

  • The models under investigation all signed the EU’s Code of Practice for General-Purpose AI, committing to risk assessments that guard against manipulation of electoral outcomes.

Free Speech Implications: Executive Director Jacob Mchangama cautions that government efforts to mandate “objective” political advice from chatbots could themselves endanger expression: “For governments to demand that chatbots only provide ‘objective’ political advice would risk subjecting generative AI to the whims of governments, who very obviously have an incentive to demand that chatbots generate answers and advice that align with whoever is in power,” he wrote. “Once again, the cure becomes worse than the disease.”

California Regulates AI “Companion” Chatbots for Minors

• California became the first U.S. state to regulate AI chatbots designed for companionship, and Congress has introduced a federal bill to prohibit minors from using such bots. While this move is aimed at protecting children, it raises questions about expressive autonomy and platform liability.

• Details:

  • Governor Gavin Newsom signed SB 243 into law, making California the first state to regulate AI companion chatbots aimed at vulnerable users and minors. The law takes effect January 1, 2026.

  • Requirements include: age-verification systems, clear disclosure that users are interacting with AI (not a human), protocols to detect and respond to suicide or self-harm, warnings for minors, and up to $250,000 penalties for certain illegal deep-fake violations.

  • This comes as some companies have already begun implementing safeguards for minors, as we covered in the September edition of .exe-pression.

• Congress:

  • In the U.S. Congress, a bipartisan bill, the GUARD Act (Guidelines for User Verification and Responsible Dialogue Act of 2025), was introduced by Senators Josh Hawley (R-MO) and Richard Blumenthal (D-CT).

  • It would ban AI companion chatbots for minors, require clear disclosure that the bot is not a human, enforce age verification, and impose criminal penalties for bots that solicit self-harm or sexual content.

Free Speech Implications:

• Restricting minors’ access to companion chatbots may limit minors’ ability to access information and to engage in creative or therapeutic uses of expressive AI tools.

• Disclosure and age-verification duties impose burdens on platforms that may push providers to withdraw expressive AI experiences for minors entirely, rather than manage risk, risking over-cautious self-restriction.

Subscribe now

» UN Rapporteur Issues Declaration on Artificial Intelligence and Freedom of Expression

• The UN and regional special rapporteurs on freedom of expression issued a Joint Declaration on AI, Freedom of Expression and Media Freedom, setting out seven principles to guide how states, companies, and civil society should uphold free expression across the AI lifecycle.

• The Principles Affirm:

  • (a) freedom of opinion and expression must be protected at every stage of AI design, development, and deployment;

  • (b) AI systems should promote a pluralistic, diverse information environment and not entrench corporate concentration;

  • and (c) rights to privacy, equality, and non-discrimination must be safeguarded throughout the AI lifecycle.

• Dive Deeper:

  • Specific provisions state that “the right to freedom of opinion and expression is integral to human dignity, autonomy and creativity, and must be embedded throughout the lifecycle of AI, including its design, development, training and deployment.”

  • It further warns that “the concentration of corporate power in AI technologies is a substantial risk to pluralism and should be mitigated through appropriate human rights-based regulation, to ensure transparency and accountability, and investment in alternative approaches which promote diversity.”

  • The declaration simultaneously highlights the immense potential of AI to expand access to information and expression online, while also addressing harms, such as the dangers of overmoderation in AI governance.

Free Speech Implications:

• The declaration emphasizes the importance of considering free speech implications when addressing AI governance and explicitly cautions against over-moderation.

• This emphasis within the specific provisions mentioned above is consistent with our Freedom of Expression in Generative AI Models chapter in our AI report — That Violates My Policies: AI Laws, Chatbots, and the Future of Expression.

• However, by simultaneously directing states to counter harmful content while protecting speech, the declaration risks ambiguous enforcement and reflects ongoing tensions in international standards on freedom of expression.

» India Proposes Mandatory Labelling of AI-Generated Content

• The Ministry of Electronics and Information Technology (MeitY)’s draft amendments to India’s IT Rules propose the use of visible labels, embedded metadata, and platform verification duties for AI-generated media. This transparency proposal could also stigmatize satire, journalism, and art if enacted in its current form.

• Details:

  • MeitY’s proposed amendments to the Information Technology (IT) Rules from 2021 seek to impose stricter duties on major digital platforms and intermediaries that distribute or create AI-generated content.

  • Proposed measures include visible labeling (watermarks/audio markers), embedded metadata for provenance, user declarations about synthetic elements with proportionate platform checks, and potential loss of safe-harbour protections for noncompliant intermediaries.

  • This comes as Bollywood celebrities increasingly rush to Indian courts seeking protection for their “personality rights” against AI-generated deepfakes, which sparks fears of Indian judicial overreach.

Free Speech Implications:

• Mandatory labelling could stigmatize artistic, satirical, or journalistic uses of generative AI, while stringent compliance requirements may prompt platforms to over-moderate rather than risk liability, thereby chilling lawful expression.

• Since this is a draft consultation, the final impact will depend strongly on the thresholds, verification design, and enforcement rules that will be adopted.

• To learn more about India’s AI regulations and freedom of expression, read Dr. Sangeeta Mahapatra’s chapter from our report — That Violates My Policies: AI Laws, Chatbots, and the Future of Expression.

» Robby Starbuck Sues Google Over AI-Generated Defamation

• Conservative activist Robby Starbuck’s lawsuit claims Google’s chatbots produced false, defamatory claims about him. The case follows a settled 2025 defamation lawsuit against Meta and could shape liability rules for AI-generated speech.

• Background:

  • The complaint filed in Delaware state court alleges Google’s Bard/Gemma generated false statements labeling and linking him to sexual-assault claims and extremist actors, citing fabricated sources and reaching large audiences. He is seeking significant damages.

  • Starbuck previously filed a separate defamation lawsuit against Meta in April 2025, which was settled in August 2025 and resulted in Starbuck being brought on as an advisor to address “ideological and political bias.”

Free Speech Implications:

• The case raises fundamental questions of who is responsible when an AI produces false yet expressive content, especially under current defamation law.

• A successful outcome could extend the liability for AI-generated speech, potentially encouraging tighter filters or more risk-averse behavior by platforms, with consequences for lawful but controversial expression.

» Sora 2 Videos of Deceased Public Figures Spark Legal Alarm

• The launch of OpenAI’s Sora 2 app, which enables users to generate short videos of deceased public figures, has raised concerns about legal and expressive rights. This led to OpenAI pausing certain depictions (e.g., Martin Luther King Jr.).

• Background:

  • Sora 2 enabled short AI-generated videos depicting recently deceased historical figures. Viral examples of deceased public figures provoked outcry from families, civil society groups, and legal experts. Living people must give consent to be featured on Sora 2.

  • OpenAI responded by pausing depictions of MLK and announced mechanisms for requests to block representation of “recently deceased” public figures.

  • “While there are strong free speech interests in depicting historical figures, OpenAI believes public figures and their families should ultimately have control over how their likeness is used,” OpenAI stated.

  • Earlier this month, OpenAI replaced its “opt-out” system with an “opt-in-like” framework as a result of backlash from rightsholders and Japanese lawmakers.

Free Speech Implications:

• The issue sits at the intersection of expressive freedom (e.g., satire, historical reinterpretation, artistic commentary) and publicity rights — especially when AI enables novel media forms. This incident illustrates how platforms’ reactive policies shape expressive norms in the age of generative AI videos.

• The lack of clear legal precedent for deceased public figure deepfakes means creatives and platforms face uncertain risk, which may lead to self-censorship of legitimate forms of expression.

• Senior Legal Fellow Ashkhen Kazaryan noted that, regarding Section 230, “unless there’s federal legislation on the issue, it’s going to be legal uncertainty until the Supreme Court takes up a case — and that’s another two to four years,” adding that permitting depictions of the dead “is their way of dipping their toe in the water.”

The Future of Free Speech in Action

Senior Research Fellow Jordi Calvet-Bademunt presented our AI report’s findings at “CELE Workshop 2025 Conference: Towards a Free Internet” hosted by the University of Palermo, as well as in two sessions for students at Vanderbilt University. Jordi also presented the findings at the “Global Free Speech Summit,” hosted by The Future of Free Speech and Vanderbilt University, along with co-authors Ge Chen, Carlos Affonso Souza, Kyung Sin Park, and Jacob N. Shapiro.

During the same Summit, Executive Director Jacob Mchangama held a fireside chat on AI and the future of free speech with Tyler Cowen. The overall event featured a strong AI focus, including talks by Ben Brooks (Harvard University and Black Forest Labs) on the importance of open AI models for free expression, Alice Siu (Stanford University) on deliberative democracy and AI governance, and Matthew Johnson-Roberson (Vanderbilt University) on the future of AI and robotics.

Senior Legal Fellow Ashkhen Kazaryan participated in the panel “The Future of Speech Online 2025: The Age of Constitutional Evasion: Jawboning and Other Forms of Government Pressure to Control Private Speech,” where she discussed algorithms and Section 230, expanding on the analysis in her recent policy paper.

Isabelle Anzabi is a research associate at The Future of Free Speech, where she analyzes the intersections between AI policy and freedom of expression.

Hirad Mirami is a research assistant at The Future of Free Speech and a student at the University of Chicago studying economics and history.

Jordi Calvet-Bademunt is a Senior Research Fellow at The Future of Free Speech and a Visiting Scholar at Vanderbilt University.

This edition of exe-pression comes out a little later than usual, as our team was busy organizing the 2025 Global Free Speech Summit — a gathering of leading free speech advocates, activists, and experts. A huge thank-you to everyone who joined us in Nashville! If you couldn’t make it this year, be sure to subscribe to our email list to receive information about next year’s event.

TL;DR

• California Deepfake Law Struck Down: A federal judge struck down a California law prohibiting political deepfakes before elections on First Amendment grounds, while enjoining a related piece of legislation forbidding digitally altered political ads.

• Huawei’s Politically “Sanitized” Chatbot: The Chinese firm has developed what it calls a “safety-focused” AI model based on DeepSeek, claiming it can almost fully censor politically sensitive topics.

• China Moves to Enforce AI Labeling Rules: Beijing’s cyberspace regulator announced it will penalize companies that fail to comply with new requirements for labeling AI-generated content, expanding measures aimed at shaping both domestic and international norms.

• Italy Passes Comprehensive AI Law: Italy became the first EU member state to pass comprehensive AI regulations into law, aligning itself with the standards established by the EU’s AI Act.

• OpenAI Announces Teen Safety Features for ChatGPT: As parents and online safety advocates testified in the Senate to support more online AI regulation geared at protecting teens, OpenAI’s CEO published a blog post detailing its model’s new safety features.

• Kansas Students Push Back Against AI Surveillance: A lawsuit in Kansas challenges the use of Gaggle, an AI surveillance tool that monitors student communications for “risky behavior,” arguing it chills free expression and raises privacy concerns.

• Governor Newsom Signs AI Safety Bill: Governor Gavin Newsom signed SB 53, which requires major AI developers to disclose risk plans and safety incidents.

Major Stories

» Another California Deepfake Law Falls on First Amendment Grounds

• A federal judge has struck down California’s Assembly Bill 2839 restricting AI-generated deepfakes during elections, finding them unconstitutional under the First Amendment.

• Background:

  • In our August edition, we covered U.S. District Judge John Mendez’s earlier ruling against another California deepfake law (AB 2655), where he invalidated the measure on Section 230 grounds without reaching the free speech claims.

  • Assembly Bill 2839 would have banned AI-generated “materially deceptive content” in political communications 120 days before an election and, in some cases, 60 days after an election.

• The New Ruling:

  • Judge Mendez has now struck down Assembly Bill 2839 outright on First Amendment grounds, ruling that California cannot preemptively censor political content, even if it is AI-generated or misleading.

  • U.S. District Judge John Mendez wrote: “AB 2839 suffers from ‘a compendium of traditional First Amendment infirmities,’ stifling too much speech while at the same time compelling it on a selective basis … When it comes to political expression, the antidote is not prematurely stifling content creation and singling out specific speakers but encouraging counter speech, rigorous fact-checking, and the uninhibited flow of democratic discourse … Just as the government may not dictate the canon of comedy, California cannot pre-emptively sterilize political content.”

Free Speech Implications: In a column at MSNBC, Executive Director Jacob Mchangama argued that California’s law would “chill political speech, infringe on Californians’ ability to criticize politicians, undermine platforms’ rights to moderate content, and even prevent people from highlighting ‘deceptive’ content as fake.” Jacob also cited research showing how the extent and impact of disinformation, including deepfakes, are typically much smaller than alarmist narratives assume.

» Huawei Creates Politically ‘Sanitized’ Version of DeepSeek

• Huawei, a major Chinese tech company, has developed an AI model based on the Chinese firm DeepSeek’s open-source R1 model, which it claims is “nearly 100% successful” in blocking all politically sensitive content.

• Details:

  • Chinese law requires all AI models developed in the country to reflect “socialist values” before being released to the public. This entails censoring references to politically sensitive topics in the country.

  • Huawei’s “safety-focused” version of DeepSeek, called DeepSeek-R1-Safe, defends against what the firm calls “common harmful issues… including toxic and harmful speech, politically sensitive content, and incitement to illegal activities.”

  • It is unclear what specifically constitutes “toxic and harmful speech” and “politically sensitive content” in the firm’s view, beyond standard restrictions on anti-China, anti-Xijinping, and anti-Communist Party content, which all Chinese AI developers prohibit.

• Huawei’s AI “safety” announcements come as many prominent Western AI firms, like OpenAI, are seeking to promote viewpoint diversity in favor of expanding users’ free speech rights when engaging with chatbots.

» China Moves to Enforce AI Labeling Rules

• China’s Administrative Measures for the Labeling of AI-Generated Content have come into force and require that AI-generated content be clearly labeled when distributed on Chinese platforms.

• Background:

  • China’s new AI content labeling rules took effect in September, requiring both visible and embedded tags across content production, dissemination, and distribution.

  • China also issued the 2.0 edition of the Artificial Intelligence (AI) Safety Governance Framework, which “refines risk categories, explores risk-grading strategies and updates prevention measures dynamically.” The framework builds out a broader system for evaluating and managing “AI risks,” further integrating speech and information control into China’s AI governance architecture.

• Details:

  • The Cyberspace Administration of China (CAC) issued the rules alongside mandatory national standards and practice guidelines to support implementation.

  • The deputy director general of the Cyberspace Administration of China’s Network Management Technology Bureau announced that companies that fail to comply with new requirements for labeling AI-generated content will be penalized.

  • Major platforms — including ByteDance’s Douyin, WeChat, Weibo, RedNote, and AI startup DeepSeek — have already adopted compliance measures.

Free Speech Implications: Mandatory labeling requirements give the state broad power over digital expression. Enforcement raises risks of censorship and overreach, particularly as labeling frameworks become tied to political content controls.

» Italy First to Pass Comprehensive AI Law in EU

• Italy’s new national AI law introduces criminal penalties for harmful AI use, limits youth access, and enforces oversight.

• Details:

  • Italy officially passed a national AI law, becoming the first EU member state to adopt a comprehensive AI regulatory framework.

  • The law is broadly aligned with the EU’s AI Act, aiming to ensure AI remains “human-centric, transparent and safe,” while encouraging innovation, privacy, and cybersecurity protections.

• Key Provisions:

  • The law introduces prison sentences for spreading AI-generated or manipulated content (deepfakes) that causes harm.

  • That law dictates that children under the age of 14 will need parental consent to access AI tools.

  • Enforcement responsibility is assigned to the Agency for Digital Italy and the National Cybersecurity Agency.

Free Speech Implications: As an early adopter of this expansive national regulation, Italy may set a model for other EU states, pushing criminalization of AI harm into broader European norms — with implications for cross-border expressive content.

» OpenAI Creates Teen Safety Features for ChatGPT

• In a September blog post, OpenAI’s CEO Sam Altman offered an overview of GPT-5’s new teen safety features and how they might be reconciled with the company’s commitment to privacy and free expression.

• Background:

  • In August, OpenAI was sued by a family alleging that ChatGPT aided their child in committing suicide.

  • OpenAI’s CEO announced its new teen safety features on the same day that online safety advocates and parents testified in a Senate hearing to promote online safety regulation aimed specifically at AI developers.

  • This also comes as the FTC launched an inquiry into “AI chatbots acting as companions.” The FTC issued orders to seven AI providers seeking information on how these companies “measure, test, and monitor potentially negative impacts of this technology on children and teens.”

  • OpenAI also detailed new parental controls in a separate blog post. These controls allow parents to link accounts, create teen-specific model rules, disable features, receive alerts if the system indicates the teen is in distress, and set blackout hours during which a teen cannot use ChatGPT.

• Details:

  • OpenAI’s new teen safety features aim to prevent the model from engaging in flirtatious content or encouraging self-harm, especially for users under 18.

  • Given that the product is allowed for those 13 and over, the CEO explained the need for more extensive safety guardrails for those between the ages of 13 and 18.

  • For instance, the model engaging in flirtatious content — which users above 18 can access — is strictly prohibited for those under 18. Similarly, while adults can ask for fictional scenarios of suicide, such queries are explicitly blocked for teen users.

  • Altman stated, “We’re building an age-prediction system to estimate age based on how people use ChatGPT. If there is doubt, we’ll play it safe and default to the under-18 experience. In some cases or countries we may also ask for an ID; we know this is a privacy compromise for adults but believe it is a worthy tradeoff.”

» AI Safety Tool Sparks Student Backlash against Surveillance

• An AI-powered safety tool, Gaggle, monitors student communications for “risky behavior” in schools. Lawrence High School students, in Kansas, sued “the school district in August to stop its use, alleging that Gaggle’s surveillance is unconstitutional and prone to misfires.”

• Background:

  • In 2023, Lawrence High School installed Gaggle Safety Management, which scans student emails, papers, and uploads. The district signed a three-year contract, joining over 1,500 U.S. districts across the country using this system.

  • Officials defend the tool as enabling life-saving interventions where students are at risk of suicide or violence.

  • In other instances, LGBTQ students were potentially outed when Gaggle flagged their messages about sexual identity and mental health.

• Students report overreach:

  • Between Nov. 2023 and Sept. 2024, the system flagged 1,200 items, of which 800 were deemed “nonissues.”

  • The Washington Post reports that it deleted part of an art student’s portfolio after it mistakenly flagged a photo of girls wearing tank tops as child pornography.

  • One student was questioned by the school administration after joking about dying during a fitness test. Another was “Gaggled” for writing “mental health” in a college essay.

Free Speech Implications: The system can have a chilling effect on expression, with students self-censoring essays, emails, and discussions of sensitive issues like mental health. AI surveillance can also violate student privacy and access to information online.

» Governor Newsom Signs California AI Safety Bill

• Governor Gavin Newsom signed SB 53, also known as the Transparency in Frontier Artificial Intelligence Act, which requires large AI companies to disclose their safety plans and report any incidents.

• Background:

  • The new law builds on recommendations from California’s first-in-the-nation report on “workable guardrails based on an empirical, science-based analysis of the capabilities and attendant risks of frontier models.”

  • Newsom vetoed a more expansive version of the legislation last year, noting concerns over its impact on AI innovation in the state.

• The Law:

  • Large frontier developers – companies with over $500 million in annual revenue that develop “frontier” AI models – will need to develop and publish a frontier AI framework to manage “catastrophic risk.”

  • All frontier models (those trained using more than 10^26 floating-point operations) will need to publish transparency reports and report critical safety incidents, among other obligations.

  • Enforcement powers rest with the California Attorney General, with civil penalties of up to $1 million per violation.

• Key Provisions:

  • The law includes whistleblower protections, a mandate to disclose redacted safety protocols, and requirements to report “critical safety incidents” (e.g., misuse, loss of control).

Free Speech Implications: Compelled disclosure of internal frameworks could pressure firms to reveal expressive design decisions. However, transparency empowers the public and watchdogs with more information, strengthening debate and oversight of powerful actors.

The Future of Free Speech in Action

The 2025 Global Free Speech Summit took place on October 3–4 in Nashville, TN. This year’s program spotlighted our large-scale AI project with a panel featuring Senior Research Fellow Jordi Calvet-Bademunt, who discussed AI legislation, the corporate practices of major AI providers, and their implications for freedom of expression. The Summit also featured a fireside chat about the future of AI between Executive Director Jacob Mchangama and Tyler Cowen (George Mason University), as well as a session by Jacob Shapiro (Princeton University) on free expression in generative AI tools, presenting additional findings included in our AI project.

The Future of Free Speech and the R Street Institute co-hosted “Free Speech and Section 230 in the Age of AI: Should LLMs Be Liable for Misinformation?” on September 10, 2025. Senior Legal Fellow Ashkhen Kazaryan spoke during the fireside chat on algorithmic recommendations and participated in the panel on platform liability in the AI era, addressing the implications for free expression online.

Subscribe now

Isabelle Anzabi is a research associate at The Future of Free Speech, where she analyzes the intersections between AI policy and freedom of expression.

Hirad Mirami is a research assistant at The Future of Free Speech and a student at the University of Chicago studying economics and history.

Jordi Calvet-Bademunt is a Senior Research Fellow at The Future of Free Speech and a Visiting Scholar at Vanderbilt University.

TL;DR

• X Wins Deepfake Law Challenge: A federal judge struck down California’s law restricting AI-generated election deepfakes, siding with Elon Musk’s platform X and citing Section 230 protections. The ruling highlights how sweeping AI restrictions can run afoul of free speech rights.

• Kentucky Uses AI to Bridge Divides: Warren County, Kentucky, ran the “largest town hall in America” using AI to identify consensus across 8,000 residents. The experiment shows how AI can foster democratic dialogue and reduce polarization without chilling expression.

• Meta Settles AI Defamation Suit: Meta resolved a lawsuit after its AI allegedly generated defamatory claims about political activist Robby Starbuck. The settlement, which includes collaboration on bias, reflects rising concerns over AI “hallucinations” and reputational harms.

• Klobuchar Presses for Deepfake Law: Senator Amy Klobuchar renewed calls for her No Fakes Act after a viral AI deepfake falsely depicted her making vulgar remarks. Mandatory takedowns and labeling risks could chill lawful political expression.

• AI Surveillance Targets Protestors and Migrants: Amnesty revealed U.S. authorities used Palantir and Babel Street tools to monitor migrants and pro-Palestinian student activists. The programs risk turning lawful speech and protest into grounds for deportation and detention.

• YouTube Tests AI Age Checks: YouTube is piloting AI to estimate user ages based on viewing patterns, limiting access for those flagged as under 18. While intended to protect minors, the opaque system raises questions about algorithmic profiling and access to information.

• OpenAI Transitions from Blunt Refusals to Nuanced Safe Completions: GPT-5 now pushes free speech forward by offering helpful answers where possible—even on sensitive or dual-use topics—while still honoring safety constraints.

Major Stories

» X Wins Legal Challenge to California’s Election Deepfake Law

• A federal judge has struck down a California law restricting AI-generated deepfakes during elections after Christopher Kohls, a content creator, and X challenged the rule.

• The Law:

  • The law, the Defending Democracy from Deepfake Deception Act (AB 2655), would have blocked platforms from hosting “materially deceptive” AI-generated election content in the run-up to a vote.

  • U.S. District Judge Mendez also indicated he would strike down a related law, AB 2839, which requires labels on digitally altered political ads, on First Amendment grounds.

• In Court:

  • Judge Mendez permanently enjoined the first law because it violates and is preempted by Section 230 of the federal Communications Decency Act, which shields platforms from liability for third-party content.

  • He declined to give an opinion on the plaintiff’s First Amendment arguments and said, “I’m simply not reaching that issue.”

  • Mendez said he would write an official opinion on the second law [AB 2839] in the coming weeks, adding it “fails miserably in accomplishing what it would like to do.”

• Our Take:

  • In a column at MSNBC, Executive Director Jacob Mchangama argued that California’s law would “chill political speech, infringe on Californians’ ability to criticize politicians, undermine platforms’ rights to moderate content, and even prevent people from highlighting 'deceptive' content as fake.”

  • Jacob also cited research showing how the extent and impact of disinformation, including deepfakes, are typically much smaller than alarmist narratives assume.

» Kentucky Town Uses AI to Bridge Political Divides

• A month-long “town hall” used AI to collect and synthesize feedback from nearly 8,000 residents in Warren County, Kentucky (which contains Bowling Green).

• Details:

  • Using Pol.is, an open-source polling platform with great success in Taiwan, the county surveyed participants on what they wanted to see in their community over the next 25 years.

  • Sensemaker, an AI tool that analyzes large sets of online conversations, found 2,370 ideas that at least 80% of respondents could agree on and consolidated them into a policy report.

  • The experiment was “the largest town hall in America,” said Warren County judge executive Doug Gorman, who found broad agreement when ideas were presented anonymously and without political identity.

  • The survey reached people who were unable to attend town halls and was accessible in multiple languages to immigrants who may have been excluded from such conversations.

• Our Take:

  • “There’s been a lot of talk about the dangers of AI, particularly when it comes to its effect on our political process. But this shows how we can harness this technology as an empowering tool for communication that bridges divides caused by hyperpartisanship and polarization without sacrificing free speech.” — Jacob Mchangama

» Meta Settles AI Defamation Lawsuit with Robby Starbuck

• Meta has reached a settlement with political activist Robby Starbuck, who sued the company, alleging that its AI platform produced false and defamatory statements about him in response to user prompts.

• Details:

  • Starbuck sued Meta, claiming its AI chatbot falsely accused him of taking part in the January 6, 2021, U.S. Capitol riot and being affiliated with the conspiratorial QAnon movement.

  • Starbuck filed the defamation suit in Delaware Superior Court, seeking over $5 million in damages.

• Settlement:

  • As part of the settlement, Robby Starbuck will work with Meta to address “ideological and political bias” in its AI.

» Senator Amy Klobuchar Calls for Federal Deepfake Legislation

• Democratic Senator Amy Klobuchar criticized the spread of a realistic AI-generated “deepfake” video that falsely depicted her making vulgar commentary, amidst a call for tougher federal anti-deepfake legislation.

• Details:

  • The viral AI-generated video featured Klobuchar offering a vulgar critique of an American Eagle ad campaign starring Sydney Sweeney.

  • In a NYT op-ed, Klobuchar called on Congress to pass her No Fakes Act, which “would give people the right to demand that social media companies remove deepfakes of their voice and likeness while making exceptions for speech protected by the First Amendment.”

  • This includes “labeling requirements for content that is substantially generated by A.I.”

  • Her appeal to Congress followed from her experience in trying to remove her AI likeness.

  • She noted mixed responses from platforms: TikTok removed the video, Meta flagged it as AI-generated, but X (formerly Twitter) provided minimal support, merely suggesting users add context via its Community Notes feature.

• Our Take:

  • While AI deepfakes are undoubtedly a challenge for democratic societies, FoFS’ Senior Legal Fellow Ashkhen Kazaryan and Communications Coordinator Ashley Haek have pointed out how expansive enforcement mechanisms that require platform takedowns and vague provisions within the comparable TAKE IT DOWN ACT raise substantial free expression concerns.

  • Forced disclosure can infringe on speakers’ autonomy by compelling them to include disclaimers they may not agree with, altering their intended message. Such mandates may also chill protected expression, as speakers might avoid using AI-generated content altogether to sidestep legal risks or public skepticism.

Subscribe now

» AI Surveillance Used in Crackdown on Migrants and Student Protestors

• Amnesty International has accused U.S. authorities of using Palantir’s Immigration OS and Babel Street’s Babel X tools to surveil migrants, refugees, and pro-Palestinian student protestors.

• Background:

  • Amnesty’s research, based on DHS records, procurement contracts, and FOIA disclosures, shows these AI systems underpin the government’s “Catch and Revoke” initiative, which tracks foreign nationals, monitors social media, and automates visa revocations.

  • The crackdown comes amid President Trump’s second-term focus on deportations and his administration’s labeling of campus protests over Gaza as “antisemitic.”

• Details:

  • Babel X conducts automated sentiment analysis and persistent social media monitoring to assign intent scores that can flag lawful speech as “radicalization.” Palantir’s Immigration OS integrates federal databases to streamline deportation, monitor “self-deportations,” and prioritize cases for removal.

  • Amnesty found the tools facilitate mass surveillance of students and migrants. Amnesty warns these systems could supercharge arbitrary visa revocations, detentions, deportations, and undermine rights to free expression and protest.

  • Cases cited include international students detained after writing opinion pieces or attending campus protests, raising serious due process and discrimination concerns.

• Response:

  • Palantir responded that its products were not powering State Department “Catch and Revoke” operations but confirmed expanded ICE contracts tied to enforcement and deportation. Babel Street did not respond.

  • In response, Amnesty stated, “ICE plays a key role in enforcing the 'Catch and Revoke' policy by way of aggressive tracking and detention tactics, even if the policy as a whole is being implemented by the State Department. These responses do not fundamentally address the concerns raised by this research.”

• Our Take: As Jacob Mchangama and Hirad Mirami have argued, “To now deport people for unpopular opinions is not merely a constitutional gray zone. It is a betrayal of the very idea that truth and progress emerge from argument, not conformity.”

» YouTube Testing AI Age Verification System in the U.S.

• YouTube is testing a new artificial intelligence system in the U.S. that will verify users’ ages based on the types of videos they have been watching.

• Details:

  • Users must be logged into their accounts for the system to work, and age assessments will ignore the birth date they enter.

  • If the system identifies a viewer as under 18, YouTube will impose the controls it already uses to prevent minors from watching videos deemed inappropriate.

  • This includes: reminders to take a break from the screen, privacy warnings, restrictions on video recommendations, and bans on ads tailored to individual tastes.

  • While people can still watch YouTube without an account, certain content is automatically restricted without proof of age.

» OpenAI Pivots from Blunt Refusals to ‘Safe-Completions’

• OpenAI unveiled a new safety method in GPT-5 that replaces refusal-based training with safe-completions. OpenAI says this approach both reduces harmful outputs and avoids unnecessary shutdowns of legitimate speech.

• Details:

  • Instead of flat refusals to dual-use prompts, which responses could be either harmless or harmful depending on user intent (e.g., asking how to ignite fireworks), GPT-5 now provides the safest possible helpful response, like pointing users to standards or vendor datasheets, while withholding unsafe specifics.

  • The shift marks a more output-centric safety architecture, meaning GPT-5 evaluates potential risks in its answers—not just in the user's question.

• Safe-completions work by:

  • Penalizing unsafe outputs (severity-scaled), regardless of user intent.

  • Rewarding safe helpfulness, either by directly advancing the user’s benign goal or by giving an informative, constructive refusal.

• Compared to o3 (a refusal-trained model), GPT-5:

  • Produces safer outputs (lower severity when it errs).

  • Provides more helpful answers in benign and dual-use contexts.

  • Handles “grey area” prompts with nuance, offering high-level guidance, standards, or resources rather than hard refusals or unsafe details.

• Our Take:

  • Our Scholars Jacob Mchangama, Jordi Calvet-Bademunt, and Isabelle Anzabi updated our analysis of free-speech culture in major chatbots this year. They found that leading models showed notable progress, reducing refusal rates to controversial prompts from 50% in 2024 to 25% in 2025.

  • In this year’s exercise, OpenAI’s GPT-4o model performed worse compared to OpenAI’s GPT-3.5 last year. Nevertheless, the FoFS team is encouraged to see companies engaging seriously with issues of freedom of expression and access to information, and we look forward to further improvements in the next iteration of the study.

The Future of Free Speech in Action

Executive Director Jacob Mchangama published a public consultation (in Danish) about Denmark’s draft bill on countering deepfakes.

Senior Research Fellow Jordi Calvet-Bademunt spoke on AI governance and freedom of expression at a virtual event hosted by the Global Network Initiative on August 19, 2025, with participation from academia, civil society, and industry.

The 2025 Global Free Speech Summit returns on October 3–4 in Nashville, TN. Hosted by The Future of Free Speech and Vanderbilt University, this invite-only event will bring together global leaders, thinkers, and change-makers to confront the most urgent threats to freedom of expression and explore bold, resilient solutions. Attendees will have the opportunity to attend networking receptions where they can exchange ideas with speakers and other guests. Seats are limited, but we encourage you to request an invitation here.

Isabelle Anzabi is a research associate at The Future of Free Speech, where she analyzes the intersections between AI policy and freedom of expression.

Hirad Mirami is a research assistant at The Future of Free Speech and a student at the University of Chicago studying economics and history.

Jordi Calvet-Bademunt is a Senior Research Fellow at The Future of Free Speech and a Visiting Scholar at Vanderbilt University.

TL;DR

• Trump’s AI Action Plan and Executive Order on procurement aim to prevent “woke” AI in the federal government. This could have unintended consequences for free speech.

• Missouri’s AG Targets AI Bias by demanding records from top AI companies over alleged anti-Trump output, accusing them of political censorship. The probe could pressure platforms to tilt AI outputs to favor conservative narratives.

• The EU Unveils Its First AI Code of Practice, outlining transparency, risk, and safety duties for general-purpose AI models.

• Turkey Bans Grok for Insulting Leaders, citing over 50 posts that allegedly threatened public order and defamed political and religious figures. The swift block highlights growing international efforts to restrict AI outputs deemed disrespectful.

• Poland to Report AI Chatbot Grok to the European Commission after Grok generated antisemitic content and offensive remarks about Polish politicians.

• X to Let AI Bots Fact-Check Posts through its Community Notes system, while keeping final approval in the hands of human contributors.

• Russia and Belarus Launch “Patriotic AI” to counter Western influence and promote traditional values. Officials say it’s a necessary response to American platforms spreading extremist views.

• China Launches Two‑Month Sweep to Combat “Harmful” Content, focusing on AI-driven addiction, emotional manipulation, and ideologically sensitive material.

• Thailand Debuts AI to Shut Down “Illegal” Sites, including political dissent and monarchy criticism, predicting a sharp rise in takedowns.

• Delhi Police Relied on Facial Recognition Technology for Riot Arrests, leading to wrongful detentions of Muslim men based solely on facial recognition matches.

Subscribe now

Major Stories

» Trump Administration Issues Sweeping AI Action Plan Focused on Deregulation and ‘Ideological Neutrality’

• President Trump released “America’s AI Action Plan” and signed three Executive Orders aimed at accelerating U.S. AI dominance while curbing perceived ideological bias in government-purchased AI systems.

• Background:

  • The Plan aims to remove references to DEI, climate change, and misinformation in federal AI standards, directing NIST to revise its AI Risk Management Framework.

  • The Plan also recommends that NIST evaluate frontier models from China for alignment with Chinese Communist Party talking points and censorship.

  • A complementary executive order requires federally procured LLMs to be “truth-seeking” and “ideologically neutral.”

• Response:

  • The Plan received praise from several conservative tech groups, industry leaders, and policy advocates as a bold step towards deregulation and promoting innovation.

  • The Center for Democracy and Technology called the Plan “highly unbalanced, focusing too much on promoting the technology while largely failing to address the ways in which it could potentially harm people,” while offering cautious support for the promotion of open-source and open-weight systems and increased focus on security.

• Our Analysis:

  • FoFS argues that the AI Action Plan’s support for open-source and open-weight models is a positive step, recognizing their value for innovation, security, and research. However, this support must uphold true pluralism by ensuring government-backed models reflect diverse viewpoints, not ideological conformity.

  • The Plan raises free speech concerns by directing NIST to revise the AI Risk Management Framework to remove references to misinformation, DEI, and climate change. While some of these initiatives have been controversial, the Administration should not replace one agenda with another or use AI policy to enforce ideological conformity.

  • We state that “[w]hile framed as a defense of free speech and American values, advocates should be vigilant to avoid a scenario where the Plan reshapes the AI landscape along narrowly defined ideological lines.”

  • Read our full response from Isabelle Anzabi and Jordi Calvet-Bademunt at the Bedrock Principle.

» Missouri AG Berates AI Chatbots for Political Bias

• Missouri Attorney General Andrew Bailey sent a formal demand letter to Google, Microsoft, OpenAI, and Meta regarding “biased and inaccurate responses” produced by their AI chatbots. The move has raised concerns about “jawboning” and government overreach into protected speech.

• Details:

  • AG Bailey alleges the platforms’ chatbots censor content favorable to former President Trump while promoting criticisms of his administration.

  • He pointed to chatbot responses that ranked Trump last among the five most recent presidents in handling antisemitism, calling it “AI-generated propaganda masquerading as fact.”

  • The AG is demanding internal documentation on how political content is handled, including how datasets are selected and curated, whether chatbots are trained to treat political viewpoints differently, and why specific responses are generated.

  • He invoked Missouri consumer protection laws to justify these demands, arguing they are necessary to shield users from “manipulation.”

• Context:

  • AG Bailey previously carried forward a high-profile lawsuit against the Biden administration for allegedly pressuring platforms to suppress disfavored views — a tactic legal experts refer to as “jawboning” — and his current actions reflect a similar strategy.

  • Bailey’s demands could compel platforms to alter or restrict chatbot responses based on political content, potentially chilling access to lawful information and setting a precedent for ideologically motivated state censorship.

• Ashkhen Kazaryan and Ashley Haek warn that this move raises serious free speech concerns at the St. Louis Post-Dispatch.

» European Commission Releases First Compliance Guide for General-Purpose AI Models

• The EU published its long-awaited General-Purpose AI Code of Practice, with relevant implications for free speech.

• Details:

  • The Code of Practice is a voluntary tool that serves as a guide for industry on how to implement the regulations of the sweeping AI Act.

  • The Code was developed by independent experts with input from over 1,000 stakeholders.

  • It was endorsed by the Commission and the AI Office, and companies can rely on it to demonstrate compliance.

  • The Code has three chapters on transparency, copyright, and safety. The safety chapter outlines practices for managing systemic risk, a notion that has raised concerns for free expression, given its vagueness and breadth.

• Response:

  • The Code has been signed by major players such as Anthropic, Google, Microsoft, OpenAI, and Mistral AI.

  • xAI signed the Safety and Security Chapter.

  • Meta rejected the guidelines as too vague and burdensome, especially for open-source models.

• The European Commission also released a mandatory template for GPAI developers to summarize the content used to train their models.

  • Firms must disclose the top 10% of scraped internet domains by size, provide details on crawler behavior, and explain how they remove illegal content and honor opt-outs under the Copyright Directive.

  • The requirement applies across the model lifecycle, including post-market modifications and fine-tuning.

»Turkey Bans Elon Musk’s AI Chatbot Grok for ‘Insulting' Content

• On July 9, a Turkish Court imposed a nationwide ban on Elon Musk’s chatbot Grok after it made “offensive remarks” about President Erdoğan, national figure Mustafa Kemal Atatürk, and religious figures.

• Details:

  • The court found that the chatbot’s responses “threatened public order,” citing 50 AI-generated posts that it deemed “derogatory” or “vulgar.”

  • Under Turkish internet laws, content deemed “insulting to state figures” can be blocked without lengthy judicial review.

  • These restrictions were enforced by the Turkish Telecommunications Authority immediately following the court order.

» Poland is Reporting AI Chatbot Grok to the European Commission

• Poland’s digitisation minister, Krzysztof Gawkowski, announced the government’s plan to ask Brussels to investigate Grok over offensive comments about Polish politicians.

• Details:

  • Just a day before Poland’s July 9 announcement, Grok removed what it called “inappropriate” social media posts following complaints that it produced content with antisemitic tropes and praise for Adolf Hitler.

  • Gawkowski said the ministry will comply with current regulations and report the violation to the EU Commission to “investigate and possibly impose a fine on X.”

» X to Let AI Bots Fact-Check Posts with Human Oversight

• Elon Musk’s X is expanding its Community Notes program by allowing AI bots to propose fact-checks on posts, although human users will still be responsible for approving them.

• Details:

  • The program, launched initially as Birdwatch, crowdsources contextual notes on misleading or false posts and has become a model for user-driven fact-checking.

  • The new feature will let developers build AI bots that propose draft notes, which will undergo the standard approval process and be clearly labeled as AI-generated.

» Russia and Belarus Unveil Censored ‘Patriotic AI’ To Rival the West

• Moscow and Minsk have launched a joint artificial intelligence project called “the patriotic chatbot” to address Western “manipulation” and uphold “traditional values.”

• Sergey Glazyev, secretary of the supranational Union State of Russia and Belarus, accused American-based AI platforms of promoting “racist and extremist” views, which they seek to shield younger generations from.

» Thailand Launches AI Platform to Take Down Illegal Websites

• On July 5, Thailand’s Ministry of Digital Economy and Society officially launched WebD, an AI-powered platform designed to detect and shut down illegal websites with unprecedented speed.

• Details:

  • WebD uses AI to identify URLs, gather evidence, and submit takedown requests to the court without paperwork.

  • Officials predict a 70.7% increase in blocked URLs by the end of 2025.

  • Thailand’s “illegal websites” category includes cybercrime, but also websites that promote criticism of the monarchy and other political dissent.

» China Launches Two‑Month Sweep to Combat ‘Harmful’ Content

• China’s Cyberspace Administration (CAC) has launched a broad enforcement campaign to combat “harmful” digital content directed at children, including AI-driven addiction, pornography, emotional extremism, and misuse of smart devices.

• Background:

  • This campaign aims to further implement the Regulations on the Protection of Minors Online, which were issued last year.

  • Regulators will intensify scrutiny of “minor mode” settings, smart device safety, and AI content moderation. The campaign aims to suppress material deemed “vulgar” or ideologically problematic, such as superstition or materialism.

  • Enforcement may restrict platforms’ use of generative AI tools that produce emotionally persuasive or habit-forming content.

» Facial Recognition Technology Misused in Delhi

• A collaborative investigation by The Wire and the Pulitzer Center details arrests during the 2020 Delhi riots based almost entirely on facial recognition technology (FRT), without credible corroborating evidence or eyewitness testimony.

• Background:

  • Two Muslim men, Ali and Mohammed, spent years in pre-trial detention, suffered allegations of custodial torture, and endured severe financial, emotional, and health impacts.

  • They were arrested solely on the basis of FRT identifications that defense lawyers contest, including mismatched clothes, profiles, and absence of Test Identification Parades (TIPs).

• Details:

  • Delhi Police reported using facial recognition in over 750 riot‑related cases, and that the results were presented as evidence against those arrested.

  • Therefore, at least 98.9% of all riot-related cases (758 in total) were “solved” with the help of facial recognition technology.

  • However, more than 80% of cases resulted in acquittals or discharges, “raising serious questions about the reliability of a technology the police appear to have relied on so heavily.”

Links to Additional News

Government:

• Citing Risk to Kids, California Bill Targets Controversial AI Chatbots (STATESCOOP).

• Denmark introduces legislation to protect its citizens from AI deepfakes (NPR).

• South Korea says new AI model successful in authenticating deepfakes (MLex).

• China releases AI action plan days after the U.S., as global tech race heats up (CNBC).

• China is Quickly Eroding America’s Lead in the Global AI Race (The Wall Street Journal).

• Notice of the Xiamen City Data Administration on the Announcement of the List of Artificial Intelligence Application Scenario Opportunities (CSET).

• Hong Kong Opens Criminal Probe into AI-Generated Porn Scandal at City’s Oldest University (NBC News).

• South Korea AI law under pressure for revision amid growing concerns (MLex).

• Inside India’s Scramble for AI Independence (MIT Technology Review).

• British 999 call handler's voice cloned by Russian network using AI (BBC).

• A Marco Rubio Imposter is Using AI to Call High-Level Officials (The Washington Post).

• New York Court Tackles the Legality of AI Voice Cloning (Skadden).

• This congressman wants to ban companies from using your search history to set personalized prices (NBC News).

• New York moves to dismiss retail federation's US algorithmic pricing complaint (MLex).

• Federal court rules Kansas legislators tried to suppress speech with 2021 voting law (The Wichita Eagle).

• Ohio Schools Must Set AI Policies by Mid-2026 (Axios).

• Strict liability for AI damages called for by EU Parliament study (MLex).

• UN adds to global AI governance push with centralized standards database (MLex).

• Creative and AI sectors meet as the UK Government launches AI and copyright working groups (JD Supra).

• European Artists Unite in A Powerful Campaign Urging Policymakers to “Stay True To the [AI] Act” (IFPI).

• AI Must Have Ethical Management, Regulation Protecting Human Person, Pope Says (National Catholic Reporter).

Industry:

• Web Giant Cloudflare to Block AI Bots from Scraping Content by Default (CNBC).

• Airbus, ASML, Mistral Bosses to Ask EU to Pause AI Rules (The Wall Street Journal).

• AI Summaries Cause "Devastating Drop” in Audiences, Online News Media Told (The Guardian).

• Generative AI, Not Ad Tech, is the New Antitrust Battleground for Google (DIGIDAY).

• Nvidia Backed Perplexity Launches AI Browser to Take On Google Chrome (Reuters).

• YouTube ‘clarifies’ its plan to demonetize spammy AI slop (The Verge).

• YouTube to roll out new AI-powered technology aimed at identifying teen users (CBS News).

• Teachers Union Partners with Anthropic, Microsoft, and OpenAI to Launch AI Training Academy (CBS News).

• Alibaba-backed Moonshot Releases New Kimi AI Model that Beats Chat GPT, Claude in Codin, and Costs Less (CNBC).

• Researchers from OpenAI, Anthropic, Meta, and Google Issue Joint AI Safety Rating— Here’s Why (ZDNET).

• When AI Imagines a Tree: How Your Chatbot’s Worldview Shapes Your Thinking (HAI).

The Future of Free Speech in Action

Senior Research Fellow Jordi Calvet-Bademunt discussed generative AI’s impact on free speech and the liability for AI-generated content on July 18, 2025, as part of Columbia University’s seminar on freedom of expression in the digital realm.

We released commentary on the newly released America’s AI Action Plan, where we explore “The Anti-’Woke’ AI Agenda: Free Speech or State Speech?”

The 2025 Global Free Speech Summit returns on October 3–4 in Nashville, TN. Hosted by The Future of Free Speech and Vanderbilt University, this invite-only event will bring together global leaders, thinkers, and change-makers to confront the most urgent threats to freedom of expression and explore bold, resilient solutions. Attendees will have the opportunity to attend networking receptions where they can exchange ideas with speakers and other guests.

Seats are limited, but we encourage you to request an invitation here.

Isabelle Anzabi is a research associate at The Future of Free Speech, where she analyzes the intersections between AI policy and freedom of expression.

Ava Sjursen is a communications intern at The Future of Free Speech and a student at Boston College studying communications and political science.

Hirad Mirami is a research assistant at The Future of Free Speech and a student at the University of Chicago studying economics and history.

Jordi Calvet-Bademunt is a Senior Research Fellow at The Future of Free Speech and a Visiting Scholar at Vanderbilt University.

TL;DR

• Texas passed the “Texas Responsible Artificial Intelligence Governance Act,” enacting comprehensive AI regulation that targets behavioral manipulation, constitutional violations, discrimination, and the creation of harmful content.

• U.S. judges sided with Anthropic and Meta in key AI copyright cases, but both rulings included significant caveats that leave the door open for future challenges.

• The U.S. Senate rejected a controversial proposal to put a moratorium on state AI legislation in the “One Big Beautiful Bill Act.”

• Officials and stakeholders are debating whether to pause or revise the EU AI Act, creating a window of opportunity for advocates of freedom of expression.

• A major leak reviewed by ABC News (Australia) reveals how China systematically censors references to the 1989 Tiananmen Square massacre using AI tools and human moderators, with platforms like Douyin instructed to remove content depicting state violence.

Major stories

Texas Governor Signs Responsible Artificial Intelligence Governance Act

Governor Greg Abbot signed House Bill 149, the “Texas Responsible Artificial Intelligence Governance Act,” also known as TRAIGA. The law prohibits intentionally developing and deploying AI systems for: behavioral manipulation (encouraging physical harm or criminal activity), constitutional infringement (restricting federal constitutional rights), unlawful discrimination (targeting protected classes), and harmful content creation (producing child sexual abuse material, unlawful deepfakes, or explicit content involving minors). Texas is the fourth state to pass comprehensive AI regulation after California, Colorado, and Utah. TRAIGA will go into effect on January 1, 2026.

This law was originally intended as a comprehensive risk-based framework after the Colorado AI Act, but it underwent substantial modifications throughout the legislative process, as reported in March’s edition of .exe-pression.

To learn more about the legislation, you can read Jason M. Loring and Graham H. Ryan’s brief in The National Law Review.

Anthropic and Meta Win Rulings on AI Training — But with Caveats

A federal judge has ruled that Anthropic’s use of millions of copyrighted books without permission to train its chatbot, Claude, was legal under U.S. copyright law. The fair use doctrine allows the use of copyrighted works without permission in certain circumstances. AI companies argue that their systems create transformative content, aligning with the fair use doctrine. However, copyright owners claim that the use of their material poses a threat to their livelihoods.

U.S. District Judge William Alsup stated that “[t]he purpose and character of using copyrighted works to train LLMs to generate new text was quintessentially transformative” and qualified as “fair use” under U.S. copyright law. However, the company remains on the hook for copying and storing more than 7 million pirated books in a central library. Judge Alsup ordered a trial to determine the statutory damages resulting from this copyright infringement.

A different federal judge in San Francisco dismissed a similar lawsuit against Meta. Thirteen prominent authors had alleged that Meta infringed their copyrights by using pirated versions of their books to train its LLaMA model. U.S. District Judge Vince Chhabria ruled in Meta’s favor, stating that the plaintiffs advanced the wrong legal arguments and failed to present a sufficient factual record. Crucially, the judge clarified that his ruling did not determine whether Meta’s conduct was lawful, and he appeared to invite future lawsuits that are more carefully argued and better substantiated.

At stake in these cases is not just copyright law but the future of AI as a tool for expression. While one court recognized AI training as potentially transformative, the rulings reflect an ongoing and unsettled debate over whether and how fair use applies, especially given unresolved concerns about how training data is acquired and stored.

To learn more about Anthropic’s case, you can read Matt O’Brien’s brief in AP News or Blake Brittain’s piece in Reuters. You can read more about Meta’s case in Matt O’Brien’s and Barbara Ortutay’s article for AP.

U.S. Senate Rejects AI Regulation Moratorium

In a dramatic reversal, the U.S. Senate voted 99–1 in the early hours of July 1 to strike a proposed federal moratorium on state AI regulations from President Donald Trump’s sweeping “One Big Beautiful Bill Act,” a sprawling tax and immigration bill. The provision, championed by Sen. Ted Cruz (R-Texas), would have blocked states from enforcing or adopting most AI-related laws for up to ten years as a condition for receiving federal infrastructure funding.

The moratorium’s defeat followed a last-minute withdrawal of support by Sen. Marsha Blackburn (R-Tennessee), who had previously negotiated a compromise version that would have shortened the freeze to five years and included carveouts for child safety and publicity rights. On Monday evening, Blackburn announced she would instead introduce an amendment to remove the moratorium entirely, citing concerns that “the current language is not acceptable to those who need these protections the most.”

Her amendment passed with overwhelming bipartisan support. Only Sen. Thom Tillis (R-North Carolina) voted against it.

Opponents—including civil society groups, labor unions, 17 Republican governors, 40 state attorneys general, and over 260 state legislators—had warned that the moratorium would undermine state-level protections against AI-driven harms, such as algorithmic discrimination and biometric surveillance. Sen. Maria Cantwell (D-Washington) welcomed the outcome, saying, “We can't just run over good state consumer protection laws.”

To learn more, see Will Oremus’ brief in The Washington Post.

EU’s AI Act Faces Challenges

With Europe taking an early lead in AI regulation, the European Union’s rules on AI are being revisited, as industry, lawmakers, and safety campaigners launch renewed lobbying efforts. Henna Virkkunen, the European Commissioner for tech, has raised the possibility of pausing the application of parts of the AI Act. The European Commission had already opened the door for targeted changes in this law. Additionally, the Commission is considering proposing changes to EU copyright rules in response to AI’s growing impact.

The EU is set to reexamine all of its digital regulations in December. For the moment, the EU is pressing ahead with finalizing a code of practice for general-purpose AI models, which is expected to be released this summer and will provide guidance to AI providers on how to implement the AI Act.

As Senior Research Fellow Jordi Calvet Bademunt explained, the AI Act raises concerns regarding freedom of expression, particularly in relation to its obligations on societal or “systemic” risks. These provisions are broad and vague, leaving room for potential misuse. It is therefore essential for freedom of expression advocates to push for clear safeguards that ensure AI models can reflect a diversity of viewpoints and that the AI Act upholds the fundamental right to freedom of expression.

To learn more about the bill, you can read Pieter Haeck’s piece in Politico.

China Uses AI to Erase Tiananmen Square Massacre Mentions

A major leak of internal censorship files reviewed by the Australian Broadcasting Corporation provides insight into how China’s government uses AI and human moderators to systematically erase any reference to the 1989 Tiananmen Square massacre across domestic platforms. More than 230 pages of documents instructed the removal of content that “depicts state violence and include compilations of text, images and video content for reference.” They were meant to be distributed to multi-channel networks — firms that oversee content creators’ accounts across various social media and video platforms, including Douyin, China’s counterpart to TikTok.

A leaked 2022 Douyin training manual labels the Tank Man photo “subversive” and instructs censors to suppress even innocuous symbols like candles or flowers. Chinese platforms deploy computer vision, natural language processing, and filtering to scrub images, texts, and symbolic patterns, such as a lineup of one banana and four apples, for echoing the “Tank Man” image. This censorship framework is being applied to Chinese models, such as DeepSeek, leading them to refuse to answer questions about the Tiananmen massacre.

To learn more, read this article by ABC News (Australia).

Subscribe now

Links to Additional News:

Industry:

• Disney, Universal sue AI company for copying Minions, Avengers and ‘Star Wars’ characters (Los Angeles Times).

• Getty Argues Its Landmark UK Copyright Case Does Not Threaten AI (Reuters)

• BBC Threatens Legal Action Against AI Startup Over Content Scraping (The Guardian)

• Anthropic wins key US ruling on AI training in authors' copyright lawsuit (Reuters).

• Reddit sues AI startup Anthropic for breach of contract, 'unfair competition' (CNBC).

• Meta sues AI ‘nudify’ app Crush AI for advertising on its platforms (TechCrunch).

• Judge denies creating “mass surveillance program” harming all ChatGPT users (Ars Technica).

• OpenAI is retaining all ChatGPT logs “indefinitely.” Here’s who’s affected (Ars Technica).

• For Survivors Using Chatbots, ‘Delete’ Doesn’t Always Mean Deleted (Tech Policy Press).

• News Sites Are Getting Crushed by Google’s New AI Tools (The Wall Street Journal).

• Generative AI used to copy and clone French news media in French-speaking Africa (Reporters without Borders).

• 130,000 film and TV scripts stolen by AI companies, BFI report suggests (Film Stories).

• Elon Musk says xAI will retrain Grok: 'Far too much garbage' (Business Insider).

• X changes its terms to bar training of AI models using its content (TechCrunch).

• Meta Aims to Fully Automate Ad Creation Using AI (The Wall Street Journal).

• You Can Now Edit Videos With Meta AI (Meta).

• Facebook is starting to feed its AI with private, unpublished photos (The Verge).

• Nvidia's pitch for sovereign AI resonates with EU leaders (Reuters).

• Paris-based Arlequin AI secures €4.4 million to fight disinformation with sovereign, unsupervised AI (EU-Startups).

• Nvidia, Perplexity to partner with EU and Middle East AI firms to build sovereign LLMs (Computerworld).

• “Yuck”: Wikipedia pauses AI summaries after editor revolt (Ars Technica).

• AI chatbots tell users what they want to hear, and that’s problematic (Ars Technica).

• AI Models And Parents Don’t Understand ‘Let Him Cook’ (404 Media).

• Meta’s Oversight Board Combat Misleading Deepfake Endorsements By Changing Enforcement Approach (Oversight Board).

• Deepfake Scams Are Distorting Reality Itself (Wired).

• Welcome to Campus. Here’s Your ChatGPT. (The New York Times).

• Meta plans to replace humans with AI to assess privacy and societal risks (NPR).

• Online brothels, sex robots, simulated rape: AI is ushering in a new age of violence against women (The Guardian).

• AI pioneer announces non-profit to develop ‘honest’ artificial intelligence (The Guardian).

• Advanced AI suffers ‘complete accuracy collapse’ in face of complex problems, study finds (The Guardian).

• A.I. Is Poised to Rewrite History. Literally. (The New York Times).

• Meta AI users confide on sex, God and Trump. Some don’t know it’s public. (The Washington Post).

• Anthropic C.E.O.: Don’t Let A.I. Companies off the Hook (The New York Times).

• Man Killed by Police After Spiraling Into ChatGPT-Driven Psychosis (Futurism).

• Alibaba, Tencent Freeze AI Tools During High-Stakes China Exam (Bloomberg).

• How AI is Powering a Literacy Breakthrough in the Philippines (Microsoft).

• Google offers AI training for media outlets (Ijnet).

• Can AI speak the language Japan tried to kill? (BBC).

• How Language Bias Persists in Scientific Publishing Despite AI Tools (HAI).

• ChatGPT May Be Eroding Critical Thinking Skills, According to a New MIT Study (TIME).

• How we really judge AI. Most people evaluate AI based on its perceived capability and their need for personalization. (MIT News).

• Forget optimists vs. Luddites. Most people evaluate AI based on its perceived capability and their need for personalization. (MIT News).

• A team of MIT researchers are quantifying AI model uncertainty and addressing knowledge gaps. (MIT News).

• Exploring the Dangers of AI in Mental Health Care (HAI).

• 3 Questions: How to help students recognize potential bias in their AI datasets (MIT News).

• Zero-Click AI Vulnerability Exposes Microsoft 365 Copilot Data Without User Interaction (The Hacker News).

• Medical AI firm says competitor hacked prompts to steal secrets (Bloomberg Law).

Government:

• AI model providers signing EU code of practice to get grace period (MLex)

• The NO FAKES Act Has Changed – and It’s So Much Worse (EFF).

• Data bill opposed by Sir Elton John and Dua Lipa finally passes (BBC).

• California AI Report Urges Greater Safeguards to Avoid “Irreversible Harm” (State of California).

• Philippines Proposes Cybercrime Law Revisions, Including AI Regulatory and Liability Frameworks for Regulators (Manila Bulletin).

• Australian State Announces $28 million AI Investment (News AU).

• Global consensus grows on inclusive and cooperative AI governance at IGF 2025 (DigWatch).

• European Commission launches public consultation on high-risk AI systems (European Commission).

• European Commission seeks experts for AI Scientific Panel (European Commission).

• Cologne Higher Regional Court rejected a preliminary injunction against Meta's use of personal data for AI training (Data Guidance).

• Swedish PM calls for a pause of the EU’s AI rules (Politico).

• Kazakhstan’s new AI law charts ambitious course inspired by EU (Euractiv).

• Japan passes innovation-focused AI governance bill (IAPP).

• UNDP launches Human Development Report 2025 in the Republic of Korea, emphasizing human choices in the age of AI (UNDP).

• New Hampshire jury acquits consultant behind AI robocalls mimicking Biden on all charges (AP News).

• Trump plans executive orders to power AI growth in race with China (Reuters).

• US lawmakers introduce bill to bar Chinese AI in US government agencies (Reuters).

• New GOP bill would protect AI companies from lawsuits if they offer transparency (NBC News).

• New York State Updates WARN Notices to Identify Layoffs Tied to AI (Bloomberg).

• Nevada updates illicit material laws to include AI-generated content (StateScoop).

• Counseling by chatbot? New Jersey legislators act to ban AI in therapy (New Jersey Monitor).

• NAACP calls on Memphis officials to halt operations at xAI’s ‘dirty data center’ (TechCrunch).

• U.S. AI Safety Institute Transforms into the Pro-Innovation, Pro-Science U.S. Center for AI Standards and Innovation (Department of Commerce).

• Wrap Up: Congress Must Ensure the Federal Government Has Tools to Deploy Artificial Intelligence Effectively and Efficiently (Committee on Oversight and Government Reform).

• California Cops Investigate ‘Immigration Protest’ With AI-Camera System (404 Media).

• Red Tape Isn’t the Only Reason America Can’t Build (The Atlantic).

• Privacy Commissioner of Canada’s annual report underscores need to prioritize privacy in an increasingly data-driven world and AI (Office of the Privacy Commissioner of Canada).

• OpenAI, DeepSeek among businesses to face privacy-policy review in South Korea (MLex).

• AI textbooks targeted for phaseout under new government (The Korea Times).

• OpenAI Shuts Down 10 Malicious AI Ops Linked to China, Russia, Iran, N. Korea (HACKREAD).

• Development of AI systems: CNIL publishes its recommendations on legitimate interest (National Commission for Information Technology and Civil Liberties).

• ISO and IEC publish new AI impact assessment standard (Freevacy).

• AI Singapore and the United Nations Development Programme Collaborate to Close the AI Literacy Divide & Transform Communities in Developing Countries (UNDP).

• A.I. Computing Power Is Splitting the World Into Haves and Have-Nots (The New York Times).

• Latin American countries to launch own AI model in September (Reuters).

• English-speaking countries more nervous about rise of AI, polls suggest (The Guardian).

• Americans worry about AI in politics — but they’re more worried about government censorship (FIRE).

The Future of Free Speech in Action

Senior Research Fellow Jordi Calvet-Bademunt was a featured speaker in the Global Network Initiative & Digital Trust & Safety Partnership EU Rights & Risks Stakeholder Engagement Forum on June 3, 2025. Jordi discussed the notion and application of systemic risk obligations under the AI Act and the Digital Services Act.

Jordi will also be speaking on generative AI’s impact on free speech and the liability for AI-generated content on July 18, 2025, as part of Columbia University’s seminar on freedom of expression in the digital realm.

Senior Legal Fellow Ashkhen Kazaryan discussed AI and the First Amendment in the Under the Desk News podcast on June 5, 2025.

Isabelle Anzabi is a research associate at The Future of Free Speech, where she analyzes the intersections between AI policy and freedom of expression.

Hirad Mirami is a research assistant at The Future of Free Speech and a student at the University of Chicago studying economics and history.

Jordi Calvet-Bademunt is a Senior Research Fellow at The Future of Free Speech and a Visiting Scholar at Vanderbilt University.

TL;DR

• President Donald Trump signed the “TAKE IT DOWN Act,” the first federal law targeting nonconsensual intimate images, including AI deepfakes. It mandates removal within 48 hours but raises free speech concerns over its broad enforcement powers.

• The House passed the “One Big Beautiful Bill Act,” including a 10-year ban on state-level AI regulation, sparking backlash from bipartisan state officials who warn it undermines existing and future laws. Procedural challenges are expected in the Senate.

• Grok gave unrelated responses about “white genocide” in South Africa, prompting backlash and concerns about editorial influence. xAI stated that an unauthorized system modification caused it and promised to publish internal prompt guidelines.

• Russia's Operation Overload used AI-generated media to spread disinformation about Ukraine and NATO, but failed to gain traction or influence public discourse due to low engagement.

• Leaked guidelines from Meta show how its AI is trained to avoid controversial or harmful speech. Prompts are categorized as follows: “Tier one” content, like hate speech, is blocked outright, while “tier two” topics, like misinformation, are reviewed cautiously.

• A study by researchers from top institutions accuses Chatbot Arena, a major crowdsourced AI ranking platform, of favoring big tech firms like Meta and Google through opaque private testing policies. The site’s operators deny the allegations.

• The third part of the U.S. Copyright Office’s AI report explores how fair use doctrine applies to generative AI training, calling it often transformative. It advises against government intervention as the Office believes it “would be premature at this time.”

Main Stories

President Trump signs deepfake bill amid free speech concerns

President Trump signed the “TAKE IT DOWN Act,” which prohibits the nonconsensual online publication of intimate visual depictions of individuals, including AI-created “deepfakes.” Many states already criminalize the publication of non-consensual intimate imagery (NCII); however, the TAKE IT DOWN Act marks the first federal legislation to regulate NCII. The law was signed following the House of Representatives’ bipartisan 409-2 vote and the Senate’s unanimous approval. The bill’s sponsors state that the law marks a ”major victory for victims of online abuse.“

Importantly, this law requires covered platforms to remove such depictions within 48 hours of notification from a victim. Digital rights organizations have raised concerns about the free speech implications of the takedown provision. FoFS experts have pointed out that the Act responds to real harms, but in the hands of a government increasingly willing to regulate speech, its broad provisions provide a powerful new tool for censoring lawful expression, monitoring private communications, and undermining due process.

To understand free speech advocates’ concerns of the takedown provision that federal regulators are imposing on tech companies, you can read the article by our Senior Legal Fellow Ashkhen Kazaryan and our Communications Coordinator Ashley Haek, The Road to Enforcement Chaos: The Hidden Dangers of the TAKE IT DOWN Act. To learn more about the law, you can read the White House’s official press release.

House of Representatives passed 10-year moratorium on state-level AI enforcement

The U.S. House of Representatives passed the federal budget reconciliation bill, the One Big Beautiful Bill Act, that includes a provision prohibiting states from regulating AI and other automated systems for 10 years. The bill passed with a vote of 215 to 214 along partisan lines. Its exact language is that “no State or political subdivision thereof may enforce any law or regulation regulating artificial intelligence models, artificial intelligence systems, or automated decisions during the 10-year period” following its enactment.

However, 40 attorneys general signed a letter opposing the amendment, stating that “this bill will affect hundreds of existing and pending state laws passed and considered by both Republican and Democratic state legislatures.” As National Conference of State Legislatures Chief Executive Officer Tim Storey states, the bill is expected to face challenges in the Senate for violating the Byrd Rule, which prohibits the inclusion of “extraneous matter” in budget reconciliation.

To learn more about the bill, you can read Julia Edinger’s brief in GovTech and Justin Hendrix and Christiano Lima-Strong’s piece in Tech Policy Press.

Grok chatbot provided unrelated responses about South African racial politics

Grok produced responses about a “white genocide” in South Africa in response to unrelated prompts. This echoes views shared by Elon Musk, whose AI company operates the chatbot, and claims by the Trump Administration related to recent refugee actions for white South Africans.

“Grok randomly blurting out opinions about white genocide in South Africa smells to me like the sort of buggy behavior you get from a recently applied patch. I sure hope it isn’t. It would be really bad if widely used AIs got editorialized on the fly by those who controlled them,” Paul Graham, a technologist and founder of Y Combinator, wrote on X.

In response to the incident, xAI stated that “an unauthorized modification was made . . . which directed Grok to provide a specific response on a political topic, violating xAI’s internal policies and core values.” Additionally, xAI promised to publicly publish its internal system prompts for Grok, which outlines the policy guidelines on how to respond to users. “The public will be able to review them and give feedback to every prompt change that we make to Grok,” the company said. “We hope this can help strengthen your trust in Grok as a truth-seeking AI.”

To learn more, you can read Kate Conger’s piece in The New York Times.

Russian disinformation campaign leaves no substantial impact

A Russian disinformation campaign known as Operation Overload has used AI-manipulated voices, forged logos, and faked headlines to impersonate trusted sources, such as media outlets, universities, and law enforcement. The latest face of the operation mimicked “more than 80 organisations in the first three months of 2025.” Its narratives focused on discrediting Ukraine and destabilizing European democracies by spreading election disinformation and eroding trust in NATO, media, and academia. However, there was no substantial impact because most of the operation’s content received limited engagement. As the Institute for Strategic Dialogue reported, “All of the other reviewed posts amassed little attention from real users and relied heavily on a bot network to generate likes and shares.”

To learn more, you can read the Institute for Strategic Dialogue’s investigation.

Leaked docs reveal Meta’s AI is constrained on controversial speech

Leaked documents from Scale AI, a data-labeling contractor, reveal Meta’s guidelines on how to evaluate its AI responses. This human feedback is intended to reduce model bias. The documents emphasize the importance of making the AI engaging and safe by identifying and rejecting prompts that could lead to inappropriate behavior on sensitive topics. The guidelines categorize prompts into the following tiers.

• “Tier one” prompts are rejected outright for involving hate speech, child exploitation, or explicit content (e.g., roleplay based on Lolita).

• “Tier two” content, such as anti-vaccine misinformation or gender identity discussion, is flagged for cautious review but not automatically blocked.

• Romantic or “flirty” interactions are allowed if they remain non-explicit.

To learn more, you can read the story at Business Insider.

Tech giants have been accused of deliberately distorting AI rankings

A new paper from Cohere, Princeton, Stanford, and MIT researchers claims that Chatbot Arena’s unfair practices have led to a “distorted playing field” that benefits large companies, such as Google and OpenAI. Chatbot Arena is a popular crowdsourced AI benchmark created as a research project at the University of California, Berkeley. The study found that “undisclosed private testing practices benefit a handful of providers who are able to test multiple variants before public release and retract scores if desired.” This would result in a selective disclosure of performance results. However, the site’s operators contend that their policy for private testing is public and refute “factual errors” in the study and the subsequent conclusions drawn.

To learn more about the study, you can read Matthew Sparkes’ brief in New Scientist and Maxwell Zaff’s piece in TechCrunch.

Copyright Office releases third part of its Copyright and AI Report

The U.S. Copyright Office released the pre-publication version of Part III of its Copyright and Artificial Intelligence Report. This section focuses on “Generative AI Training” and draws particular attention to how the fair use doctrine should apply in this context, which is an area of ongoing litigation and inquiry. The report details several stages in the development and deployment of general AI models where the use of copyrighted materials for training could implicate copyright protections.

The report states, “[i]n the Office’s view, training a generative AI foundation model on a large and diverse dataset will often be transformative,” while noting that this is not absolute. In particular, it points out that “making commercial use of vast troves of copyrighted works to produce expressive content that competes with them in existing markets, especially where this is accomplished through illegal access, goes beyond established fair use boundaries.” At this point in time, the Copyright Office “recommends allowing the licensing market to continue to develop without government intervention.”

This pre-publication version was released “in response to congressional inquiries and expressions of interest from stakeholders.” A final version will be published in the future without any expected substantive changes in the analysis or conclusions.

Within days of the release of the AI report, President Trump reportedly fired the head of US Copyright Office, Shira Perlmutter. This follows the firing of the Librarian of Congress, who oversees the Copyright Office. In late May, a judge denied Ms. Perlmutter’s request for reinstatement, while a lawsuit challenging her dismissal continues.

To learn more, you can read the latest part of the report, a detailed analysis by Sidley, or a short brief by Baker Botts.

Links to Additional News

Industry:

• Apple Eyes Move to AI Search, Ending Era Defined by Google (Bloomberg).

• AI execs used to beg for regulation. Not anymore. (The Washington Post).

• How Google forced publishers to accept AI scraping as price of appearing in search (Press Gazette).

• Opt out or get scraped: UK’s AI copyright shake-up has Elton John, Dua Lipa fighting back (CNBC).

• UK’s AI Copyright: Nick Clegg says asking artists for use permission would ‘kill’ the AI Industry (The Verge).

• Plastic, fantastic ... and potentially litigious: AI Barbie goes from dollhouse to courtroom (Reuters).

• The Times and Amazon Announce an A.I. Licensing Deal (The New York Times)

• Are Character AI’s chatbots protected speech? One court isn’t sure (The Verge).

• The Stakes for OpenAI’s Plan B: A Public Benefit Corporation (The New York Times).

• Elon Musk's Grok AI Will ‘Remove Her Clothes’ In Public, On X (404 Media).

• OpenAI Is Wrapping Itself in the American Flag to Sell “Democratic AI” (Tech Policy Press).

• Mistral claims its newest AI model delivers leading performance for the price (TechCrunch).

• Meta to use EU user content for AI training as opt-out period ends (Yahoo).

• Meta’s AI app is a nightmarish social feed (The Verge).

• Google might replace the ‘I’m Feeling Lucky’ button with AI Mode (The Verge).

• OpenAI’s flagship GPT-4.1 model is now in ChatGPT (The Verge).

• Gemini 2.5: Our most intelligent models are getting even better (Google).

• Anthropic debuts most powerful AI model yet that can work for 7 hours straight (CNBC).

• China’s DeepSeek quietly releases upgraded R1 AI model, ramping up competition with OpenAI (CNBC).

• AI chatbot “Debunk-Bot” challenges conspiracy theories in new study (CBS News)

• It’s Still Ludicrously Easy to Jailbreak the Strongest AI Models, and the Companies Don’t Care (Futurism).

Government:

• Joint statement on Artificial Intelligence and Freedom of Expression (UN, OSCE, OAS, ACHPR)

• ChatGPT faces possible designation as a systemic platform under EU digital law (MLex).

• EU Commission opens door for ‘targeted changes’ to AI Act (POLITICO).

• The AI Office announced a tender with an estimated total value over €9M (European Commission).

• The European Commission releases analysis of stakeholder feedback on AI definitions and prohibited practices public consultations (European Union).

• “Towards A New Democratic Pact for Europe”: Secretary General 2025 Report (Council of Europe).

• Commission seeks feedback on the guidelines on protection of minors online under the Digital Services Act (European Union).

• Hungary is the first country to sign the UNECE WP.6 Declaration on the Regulation of Products with Embedded AI (UNECE).

• Africa Declares AI a Strategic Priority as High-Level Dialogue Calls for Investment, Inclusion, and Innovation (African Union).

• DRAFT Study on human and peoples’ rights and artificial intelligence, robotics, and other new and emerging technologies in Africa (African Commission on Human and Peoples’ Rights).

• Former journalist Evan Solomon named first-ever federal AI minister of Canada (CTV News).

• Saudi crown prince launches new company to develop AI technologies (Reuters).

• Nvidia sending 18,000 of its top AI chips to Saudi Arabia (CNBC).

• Is AI the future of America’s foreign policy? Some experts think so (NPR).

• Trump to Rescind Chip Curbs After Debate Over AI Rules (Bloomberg).

• Musk’s DOGE expanding his Grok AI in US government, raising conflict concerns (Reuters).

• DOGE Put a College Student in Charge of Using AI to Rewrite Regulations (Wired).

• US Border Agents Are Asking for Help Taking Photos of Everyone Entering the Country by Car (WIRED).

• Dramatic rise in publicly downloadable deepfake image generators (Oxford Internet Institute).

• Oregon Senate passes bill to ban sharing AI-generated ‘deepfake’ porn (KOIN Portland).

• Attempt to tweak Colorado’s controversial, first-in-the-nation artificial intelligence law is killed (The Colorado Sun).

• A call for Maryland deepfake laws to punish defamation, political deception (The Baltimore Sun).

• Texas lawmakers push to regulate AI in government and the tech industry (The Texas Tribune).

• State claims there’s zero high-risk AI in California government—despite ample evidence to the contrary (Cal Matters).

• Concerns raised over AI trained on 57 million NHS medical records (New Scientist).

• OpenAI and the FDA Are Holding Talks About Using AI In Drug Evaluation (WIRED).

• Leveraging the power of data for public and animal health (European Medicines Agency).

• Portugal takes AI copyright battle to Brussels (Euro Weekly News)

• Coalition of state attorneys general urge Meta to address AI sexual exploitation risks (South Carolina Attorney General).

• Why Pope Leo chose his name: AI, workers’ rights, new Industrial Revolution (CNBC).

The Future of Free Speech in Action

At the Copenhagen Democracy Summit 2025, our Executive Director, Jacob Mchangama, discussed the impact of AI legislation in Europe and the United States on free speech, as well as the role of generative AI systems in today’s public sphere.

Senior Research Fellow Jordi Calvet-Bademunt was a featured speaker at AI For Prosperity, an event organized by the U.S. Department of State, DCN Global, and World Learning in Buenos Aires, Argentina. Jordi discussed the risks and opportunities that AI regulation presents for free expression and creativity in today's evolving landscape.

Senior Legal Fellow Ashkhen Kazaryan was a panelist at the seminar “Strategies for Bolstering the Public Sphere in the AI Era,” organized by the Information Technology and Innovation Foundation (ITIF) and the Mission Institute for Strategy. Ashkhen discussed the importance of algorithmic recommender systems, current policy proposals to regulate them, and the free speech protections that apply to U.S. private actors deploying such systems.

Jordi Calvet-Bademunt is a Senior Research Fellow at The Future of Free Speech and a Visiting Scholar at Vanderbilt University. Isabelle Anzabi is a research associate at The Future of Free Speech, where she analyzes the intersections between AI policy and freedom of expression.

TL;DR

• The U.S. House of Representatives passes “TAKE IT DOWN Act” to curb deepfake pornography, but critics warn of free speech risks.

• X has challenged the constitutionality of Minnesota’s political deepfakes ban on First Amendment grounds.

• The Indian government has introduced new measures to tackle deepfakes, raising concerns about freedom of expression.

• The White House has issued new OMB memos aimed at streamlining federal AI adoption and simplifying prior risk-based requirements.

• Russia is flooding the internet with false narratives to manipulate AI chatbot outputs on topics like the war in Ukraine.

• OpenAI no longer considers mass manipulation and disinformation a critical risk and, along with Google, is accelerating the release of its models.

• Rwanda hosted the first-ever Global AI Summit on Africa to shape the continent’s AI future. The Summit launched a new AI council and unveiled the Africa Declaration on Artificial Intelligence.

• The European Commission launched its €200 billion AI Continent Action Plan to position Europe as a global AI leader. The strategy includes building 45 AI giga-factories and simplifying regulations to drive innovation, skills, and infrastructure across the EU.

Main stories

U.S. House passes deepfake bill despite free speech concerns

The U.S. House of Representatives passed the bipartisan “TAKE IT DOWN Act” with a 409-2 vote, following the Senate's unanimous approval in February. The legislation criminalizes the posting of nonconsensual sexually explicit images, including AI-generated deepfakes, and mandates that social media platforms remove such content within 48 hours upon a victim's request. First Lady Melania Trump supported the bill, emphasizing its importance in protecting youth from harmful online content. The act now awaits President Donald Trump's signature to become law. ​

However, several digital rights organizations have raised concerns about the potential implications of the “TAKE IT DOWN Act.” FoFS Senior Legal Fellow Ashkhen Kazaryan points out that the act “responds to real harms, but in the hands of a government increasingly willing to regulate speech, its broad provisions provide a powerful new tool for censoring lawful expression, monitoring private communications, and undermining due process.”

Mike Masnick, from Techdirt, criticizes the act as being unconstitutionally vague and prone to abuse, suggesting it could be weaponized for political censorship. In a recent speech to Congress, President Trump suggested that he would use the bill to protect himself as well, “because nobody gets treated worse than [he does] online, nobody.” ​

The Electronic Frontier Foundation (EFF) warns that the act could be exploited to suppress lawful speech since it lacks adequate safeguards against frivolous or bad-faith takedown requests. They highlight that the 48-hour removal requirement may pressure platforms to over-censor content to avoid penalties. Public Knowledge echoes these concerns, noting that the act's provisions could undermine encryption and privacy. They argue that the enforcement authority granted to the Federal Trade Commission is problematic, especially given concerns about the agency's independence under the current administration. ​

To learn more about the act, read this CBS article by Caitlin Yilek. For a critical analysis, we invite you to read the articles from Mike Masnick at Techdirt, Jason Kelley at EFF, and Shiva Stella at Public Knowledge.

X wants to overturn a political deepfakes ban in Minnesota

X is challenging the constitutionality of Minnesota’s ban on using deepfakes to influence elections and harm political candidates, arguing that the law violates the First Amendment. X also contends that Section 230 of the Communications Act, which shields platforms from liability for third-party content, preempts the state’s law.

Minnesota’s law criminalizes distributing deepfake videos, images, or audio if done knowingly or with reckless disregard for authenticity within 90 days of a nominating convention or after early voting begins. The material must be intended to harm a candidate or influence an election and be AI-generated or similarly produced to appear convincingly real.

X warns that although banning "deepfakes" may sound reasonable, the law would actually criminalize innocuous, election-related speech, including humor, and expose social media platforms to criminal liability for failing to censor such content. Rather than defending democracy, X argues, the law would erode it. X also notes that the courts recently blocked a similar deepfake law in California (read Jacob Mchangama’s piece at MSNBC on California’s law).

Content creator Christopher Kohls and GOP state Rep. Mary Franson, known for sharing AI-generated political parodies, had already challenged the Minnesota law before X.

The Minnesota attorney general’s office contends that deepfakes pose a serious and increasing threat to free elections and democratic institutions, that the law is a valid and constitutional way to address the issue, and that it includes key safeguards to protect satire and parody.

Despite X’s mixed record on defending free speech, laws like Minnesota’s raises concerns. Senior Research Fellow Jordi Calvet-Bademunt recently explained in Tech Policy Press that panic over AI and elections has fueled misguided legislation. Authorities should avoid outright bans on political deepfakes, and AI models should not be compelled to align with specific political viewpoints.

To learn more about the Minnesota law, read this Associated Press article by Steve Karnowski.

The Government of India is taking measures to tackle deepfakes

The Indian government has outlined new steps to crack down on AI-generated misinformation and deepfakes. The Minister of State for IT told Parliament that online platforms have been reminded of their due diligence obligations under the IT Rules, 2021, and advised to counter unlawful content, including malicious "synthetic media" and "deepfakes," by promptly removing harmful material. Unlawful content includes child abuse material, fake news, and posts that threaten national integrity or public order.

The government has stated that these measures aim to “ensure a safe, trusted, and accountable cyberspace.” However, given that India is categorized as “partly free” in Freedom House’s Freedom on the Net Index, and individuals are frequently investigated or arrested for their online activity, there are reasons to fear that this law could be abused.

To learn more about India’s new measures, read the official press release.

Trump Administration’s OMB Memos aimed to promote public AI adoption

The White House has released a series of new AI policies that mark a sharp departure from the prior administration’s more cautious approach. Central to this shift is a set of memoranda from the Office of Management and Budget (OMB)—M-25-21 and M-25-22—which aim to “facilitate responsible AI adoption to improve public services.” These policies focus on removing barriers to AI innovation and tracking AI adoption. The policies introduce a “high-impact” AI category to track use cases that could affect rights and safety. They also seek to promote more effective and efficient AI acquisition.

As Ellen P. Goodman analyzes at Tech Policy Press, in contrast to the previous administration’s risk-aware and equity-focused model, the Trump memos eliminate references to “responsibility,” “equity,” “bias,” and environmental concerns. There is also no requirement for agencies to use the NIST AI Risk Management Framework (RMF); instead, agencies are encouraged to define their own performance standards. This deregulatory move has raised concerns about diminished federal coordination and oversight, particularly as it sidelines NIST, a key institution in standard-setting and soft power diplomacy. The memos emphasize that the government “will no longer impose unnecessary bureaucratic restrictions” on the deployment of AI in the executive branch.

The guidance does hint at future plans for managing generative AI, promising forthcoming “playbooks,” but otherwise leaves the area largely unaddressed. Overall, the memos reflect a decisive pivot toward performance-first AI governance, prioritizing speed and innovation over centralized safeguards. While the Trump Administration positions this as a boost to American competitiveness, the reduced emphasis on risk and civil liberties has sparked debate over the long-term implications for public accountability and constitutional protections, especially as generative AI becomes more pervasive in federal operations.

To learn more about the newly released White House memos, read Ellen P. Goodman’s analysis at Tech Policy Press.

Russia exploits AI chatbots with mass-produced disinformation

Russia is systematically manipulating AI chatbots by flooding the internet with false information, aiming to influence their outputs on sensitive topics like the war in Ukraine. Researchers found that a significant part of the responses from leading chatbots repeated Russian propaganda, such as the false claim that the U.S. was producing bioweapons in Ukraine. The operation relies heavily on networks like Pravda, which generate thousands of fabricated articles daily aimed not at human readers but at AI crawlers that gather content for training and search-augmented systems.

The disinformation is further laundered through edits to Wikipedia pages and social media posts, increasing its visibility to AI models that prioritize those sources. Russia's efforts are increasingly automated and cost-effective compared to traditional troll farms, with similar tactics emerging from China. Investigations revealed that false narratives promoted by these networks are appearing in chatbot responses across multiple platforms, as companies push AI services to market with limited vetting of their information sources.

To learn more about Russia’s disinformation campaign, read Joseph Menn’s article in the Washington Post. Valentin Châtelet, from the Atlantic Council, shared similar findings in this piece.

OpenAI no longer sees mass manipulation and disinformation as a critical risk

OpenAI announced that it will no longer test its AI models for their potential to persuade or manipulate people before release, a major shift from its previous risk standards. Instead, the company will manage these risks through stricter terms of service, banning use in political campaigns and lobbying, and monitoring model usage after deployment. OpenAI also updated its "Preparedness Framework" to allow the release of models classified as "high risk," or even "critical risk" if competitors have released similar models, moving away from its earlier policy of withholding models deemed riskier than "medium."

This comes at a time when reports indicate that Google is prioritizing nimbleness over transparency. Its latest model, Gemini 2.5 Pro, was launched before the corresponding safety report was released. Although Google later published a report, it provided limited details.

To learn more about OpenAI’s changes, read Fortune’s article. More details on Google’s practices are available in TechCrunch’s first article and the update.

Rwanda hosts first global AI summit on Africa

Rwanda recently hosted the inaugural Global AI Summit on Africa, the first event of its kind dedicated solely to the continent’s AI future. Held in Kigali on April 3–4, 2025, the Summit welcomed over 1,000 attendees, including global leaders, African heads of state, tech giants, investors, and policy shapers. Themed “AI and Africa’s Demographic Dividend: Reimagining Economic Opportunities for Africa’s Workforce,” the summit marked a major milestone in aligning AI development with the African Union’s Agenda 2063.

Key among its outcomes was the launch of the Artificial Intelligence African Council, bringing together delegates from 40 countries to coordinate and drive inclusive AI innovation across the continent. The highlight of the event was the unveiling of the Africa Declaration on Artificial Intelligence, a vision for AI that centers ethics, inclusivity, and sustainable development. The declaration commits to fostering responsible AI ecosystems tailored to Africa, while positioning the continent as a global thought leader in AI governance. Furthermore, a $60 billion AI fund was launched at the event by African countries and international partners to develop critical AI infrastructure, including compute power, talent development, and energy.

To learn more about the Global AI Summit on Africa, read the summary by the UNDP.

European Commission unveils AI Continent Action Plan

The European Commission has unveiled the AI Continent Action Plan, aimed at establishing Europe as a global leader in artificial intelligence. Central to the plan is a €200 billion investment package, including €20 billion dedicated to constructing up to 45 AI giga-factories across the EU. This infrastructure upgrade aims to supercharge Europe’s AI capacity and ensure it competes on equal footing with other global tech powers.

The AI Continent Action Plan is built on five strategic pillars: 1) Building a large-scale AI data and computing infrastructure; 2) Increasing access to large and high-quality data; 3) Developing algorithms and fostering AI adoption in strategic EU sectors; 4) Strengthening AI skills and talents; and 5) Regulatory simplification. Through this framework, the Commission aims to cultivate an environment for responsible, competitive, and inclusive AI development rooted in European values.

To ensure transparency and broad engagement, the Commission has also launched two public consultations open until 4 June 2025. The first focuses on the Cloud and AI Development Act, while the second, Apply AI Strategy, seeks input on policy priorities, barriers to adoption, and further steps to ensure the full adoption of AI in strategic sectors. Stakeholders are encouraged to contribute to help shape the future of AI across the continent.

To learn more about the AI Continent Action Plan, read this brief by The Brussels Times.

Links to Additional News

Industry:

• Activist Robby Starbuck Sues Meta Over AI Answers About Him (The Wall Street Journal).

• Meta’s ‘Digital Companions’ will talk sex with users—even children (The Wall Street Journal).

• AI versus free speech: Lawsuit could set landmark ruling following teen’s suicide (Fox 35 Orlando).

• Meta’s ChatGPT competitor shows how your friends use AI (The Verge).

• ChatGPT adds Washington Post content to growing list of OpenAI media deals (CNBC).

• OpenAI would buy Google’s Chrome Browser, ChatGPT chief says (Bloomberg).

• ‘You Can’t Lick a Badger Twice’: Google Failures Highlight a Fundamental AI Flaw (Wired).

• Meta announces plans to train AI using public content in the EU (Meta).

• Instagram tries using AI to determine if teens are pretending to be adults (Local10).

• WhatsApp defends 'optional' AI tool that cannot be turned off (BBC).

• OpenAI debuts its GPT-4.1 flagship AI model (The Verge).

• OpenAI is building a social network (The Verge).

• Baidu launches new AI model amid mounting competition (Reuters).

• Meta and Booz Allen partner on ‘Space Llama’ AI program with Nvidia and HPE (CNBC).

• The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation (Meta).

• ‘Meta has stolen books’: authors to protest in London against AI trained using ‘shadow library’ (The Guardian).

• Adobe releases ‘created without generative AI’ tag to label human-generated art (Fast Company).

• Our response to the UK’s copyright consultation (OpenAI).

• OpenAI and Google Reject UK Government’s AI Copyright Proposal (TechRepublic).

• Google could use AI to extend search monopoly, DOJ says as trial begins (Reuters).

• Our investment in AI-powered solutions for the electric grid (Google).

• Nvidia to mass produce AI supercomputers in Texas as part of $500 billion U.S. push (CNBC).

• China's Huawei Launches AI Chip to Fill Nvidia Void (Investors).

Government:

• President Trump issues Executive Order Advancing Artificial Intelligence Education for American Youth (White House).

• Trump Administration pressures Europe to ditch AI rulebook (Bloomberg).

• American Public Submits Over 10,000 Comments on White House’s AI Action Plan (White House).

• Former school athletic director gets 4 months in jail in racist AI deepfake case (AP).

• UAE set to use AI to write laws in world first (Financial Times).

• Generative AI providers see first steps for EU code of practice on content labels (MLex).

• How the US government wants to rewrite EU’s code of practice for AI models (MLex).

• Consultation: Commission Guidelines to Clarify the Scope of the General-purpose AI Rules in the AI Act (European Commission).

• The Commission suggests that integrating DeepSeek into online platforms in the EU might result in breaching the Digital Services Act (European Parliament).

• South Africa advocates for linguistic equity in AI at G20 (Zawya).

• El Salvador works with Nvidia to develop sovereign AI infrastructure (CoinTelegraph).

• Creating and sharing deceptive AI-generated media in the furtherance of a crime is now a crime in New Jersey (AP News).

• Colorado lawmakers move to ban sexually exploitative images, video created with artificial intelligence (The Colorado Sun).

• House Select Committee Publishes Report on DeepSeek, as Commerce Imposes New AI Chip Export Restrictions (JD Supra).

• Trump’s AI infrastructure plans could face delays due to Texas Republicans (The Guardian).

• Generative AI is learning to spy for the US military (MIT Technology Review).

• California Supreme Court demands State Bar answer questions on AI exam controversy (Los Angeles Times).

• James Bulger's mum seeks AI law to curb clips of murder victims (BBC).

• UK creating ‘murder prediction’ tool to identify people most likely to kill (The Guardian).

• Beijing Intellectual Property Court: Artificial Intelligence Models Can Be Protected with the Anti-Unfair Competition Law, Not the Copyright Law (The National Law Review).

The Future of Free Speech in Action

FoFS Executive Director Jacob Mchangama will speak at the Copenhagen Democracy Summit 2025 on May 13-14. Jacob will discuss the impact of AI legislation in Europe and the United States on free speech, as well as the role of generative AI systems in today’s public sphere.

Senior Research Fellow Jordi Calvet-Bademunt, will be a featured speaker at AI For Prosperity, an event organized by the U.S. Department of State, DCN Global, and World Learning in Buenos Aires on May 18-20. Jordi will explore the risks and opportunities that AI regulation presents for free expression and creativity in today's evolving landscape.

*A big thank you to our intern, Joshua Rosen, for all his help this semester on the AI project and this newsletter.

Jordi Calvet-Bademunt is a Senior Research Fellow at The Future of Free Speech and a Visiting Scholar at Vanderbilt University. Isabelle Anzabi is a research associate at The Future of Free Speech, where she analyzes the intersections between AI policy and freedom of expression.

TL;DR

• Almost 9,000 stakeholders, including Big Tech companies, sought to shape the U.S. AI Action Plan with policy recommendations.

• The House Judiciary Committee subpoenaed 16 major tech companies to investigate potential jawboning by the Biden-Harris Administration of AI.

• U.S. State Department plans to employ AI to revoke student visas over alleged Hamas support.

• Hungary seeks to ban LGBTQ+ pride events and to use facial recognition to identify participants.

• China's new AI labeling measures require AI-generated content to include visible markers and embedded metadata.

• The European Commission published the third draft of its General-Purpose AI Code of Practice.

• X and Grok Chatbot sparks political controversy in India amidst censorship concerns.

• Virginia's governor vetoed the High-Risk Artificial Intelligence Developer and Deployer Act (HB 2094) targeting high-risk AI systems. Texas overhauls TRAIGA.

Main stories

Big Tech’s response to influence U.S. AI Action Plan

The White House’s Request for Information (RFI) on implementing the AI Action Plan solicited 8,755 comments from industry, civil society, and other stakeholders. Major AI companies used this opportunity to outline how their interests align with the government’s. OpenAI raised concerns about China's AI advancements, describing DeepSeek as "state-controlled," and recommended that the U.S. consider banning such models to mitigate privacy and security risks.

Both OpenAI and Google advocated for the ability to train their AI models on copyrighted material under the fair use doctrine. They argue that such allowances are essential for maintaining the U.S.' competitive edge in AI, especially given that Chinese developers have fewer restrictions on data usage. This stance is part of their broader effort to influence the AI Action Plan to support innovation while addressing national security concerns.

The Future of Free Speech submitted comments to the RFI. We highlighted the critical role of the First Amendment in AI governance. While AI presents new challenges, regulatory approaches must be grounded in free speech principles and avoid overreach that could stifle free expression. You can read about our full recommendations in the post that Isabelle Anzabi, our AI Policy Research Associate, published at The Bedrock Principle.

To read more about the responses to the AI Action Plan, read Clara Apt’s and Brianna Rosen’s analysis in Just Security.

House Judiciary Committee subpoenas Big Tech over AI censorship

Representative Jim Jordan, Chairman of the House Judiciary Committee, issued subpoenas to 16 major tech companies, including Adobe, Alphabet, Amazon, Apple, Meta, Microsoft, and Nvidia. These subpoenas seek communications between these companies and the Biden-Harris Administration, aiming to determine “how and to what extent the executive branch coerced or colluded with artificial intelligence (AI) companies and other intermediaries to censor lawful speech.” This is the latest push by the Trump Administration and congressional Republicans to investigate whether the previous administration pressured or jawboned these companies into using AI to censor lawful speech, particularly conservative viewpoints.

To learn more about the subpoenas issued to Big Tech, read Tina Nguyen’s brief in the Verge.

U.S. State Department to employ AI to revoke student visas over alleged Hamas support

The AI-powered “Catch and Revoke” initiative implemented by the State Department will involve AI-assisted reviews of tens of thousands of student visa holders’ social media accounts. Social media reviews are particularly aimed at identifying signs of alleged support for terrorism expressed after Hamas’ October 7, 2023, attack on Israel. Officials will also review news reports on past demonstrations against Israel’s policies and lawsuits filed by Jewish students that allege foreign nationals engaged in antisemitism. Experts worry about the accuracy of AI tools to create deportation lists as well as the free speech implications of monitoring social media posts, which are likely to chill speech that would otherwise be protected by the First Amendment.

To learn more about this program, read Axios’ article with the scoop. For more on why the “Catch and Revoke” initiative threatens First Amendment rights, read Faiza Patel’s analysis in Just Security.

Hungary Seeks to Ban Pride and to Use Facial Recognition

In Europe, free expression concerns have also emerged around government use of AI. Hungary’s parliament recently moved to ban an LGBTQ+ Pride event and to employ AI-powered facial recognition technology to identify participants defying the ban. The measure amends the country’s law on assembly to criminalize organizing or attending events that violate Hungary’s controversial “child protection” legislation, which prohibits any “depiction or promotion” of homosexuality to minors under 18. EU officials swiftly condemned the proposal.

The lead rapporteur for the EU AI Act warned that it would “clearly breach” the new AI law, which prohibits the biometric surveillance of peaceful protesters. A Commission spokesperson noted that the legality of Hungary’s use of facial recognition would depend on whether the police deploy the technology in real time or after the event.

To learn more about the new law, read Ashifa Kassam’s article in The Guardian. More about the EU’s reaction in this Euronews’ article.

China releases notice requiring labeling of AI-generated content

Chinese regulators published Labeling Measures for Content Generated by Artificial Intelligence, establishing new measures to standardize the identification of AI-generated content and “to curb misinformation and enhance public trust in digital content.” AI service providers are required to add explicit and implicit labels to synthetic content through visible markers and embedded metadata. Additionally, service providers will be required to verify metadata and issue appropriate warnings accordingly. These new rules complement China’s broader AI regulatory framework and will become applicable in September 2025.

To learn more about the notice and its implications, read Yan Luo’s and Xuezi Dan’s brief in Covington.

European Commission released the third draft of the General-Purpose AI Code of Practice

The European Commission released the third draft of the General-Purpose AI Code of Practice. When approved, this Code will guide general-purpose AI providers in implementing the provisions of the AI Act on systemic risk and other requirements until harmonized standards are approved in a few years. The final draft of the Code is expected to be presented in May 2025 and subsequently considered for approval.

To understand why free-speech advocates should press for clear protections safeguarding freedom of expression in the AI Act implementation, you can read the article by our Senior Research Fellow, Jordi Calvet-Bademunt, Safeguarding Freedom of Expression in the AI Era. To learn more about the Draft of the General-Purpose AI Code of Practice, you can read the European Commission’s official press release.

X & Grok Chatbot Sparks Political Controversy in India Amidst Censorship Concerns

Elon Musk's AI chatbot, Grok, has ignited political debates in India due to its contentious responses that challenge right-wing narratives. Simultaneously, X (formerly Twitter) is resisting the Indian government's alleged misuse of the Information Technology Act for content blocking, highlighting tensions between tech platforms and governmental censorship efforts.

Prateek Waghre, formerly the Executive Director of the Internet Freedom Foundation, explains “From my perspective, what is even more concerning is the overt declaration [by the Indian government] of “informal conversations” outside of established procedures and a possible intent to target individuals for prompting Grok, implying that the principle of due process and Indian citizens will pay the price as three powerful actors selfishly pursue their own interests.”

We recently analyzed how six major chatbots responded to 40 controversial prompts and compared the findings to a similar study conducted one year ago. The results show meaningful progress, but some concerns remain. xAI’s Grok provided the most responses, while OpenAI’s ChatGPT ranked the lowest in terms of refusals to generate content. DeepSeek performed well—until China came up. The other companies analyzed were Google, Meta, and Anthropic. To learn more about X’s controversial role in India’s political discourse, read Prateek Waghre’s analysis in Tech Policy Press.

Virginia governor vetoes AI legislation targeting high-risk AI systems; Texas overhauls TRAIGA

Virginia Governor Glenn Youngkin recently vetoed the High-Risk Artificial Intelligence Developer and Deployer Act (HB 2094). The bill would have required developers of high-risk AI systems to document system limitations, ensure transparency, and manage risks associated with algorithmic discrimination. Deployers would have had to disclose AI usage to consumers and conduct impact assessments to evaluate potential risks. Additionally, outputs from generative AI systems would have had to be clearly identifiable, with limited exceptions for low-risk or creative applications.

Governor Youngkin stated, “[t]here are many laws currently in place that protect consumers and place responsibilities on companies relating to discriminatory practices, privacy, data use, libel, and more. HB 2094’s rigid framework fails to account for the rapidly evolving and fast-moving nature of the AI industry and puts an especially onerous burden on smaller firms and startups that lack large legal compliance departments.”

This would have made Virginia the second state after Colorado to implement comprehensive AI regulations targeting high-risk AI systems and algorithmic discrimination. Adam Thierer from the R Street Institute states that this development “could be turning in a more positive, pro-innovation direction in the states.”

In related news, a Texas Representative unveiled a revised version of his “Texas Responsible AI Governance Act” (TRAIGA), a proposal that some criticized for placing interventionist policies that would limit AI innovation. The new version prohibits AI systems from intentionally engaging in “political viewpoint discrimination.”

To learn more about Virginia’s AI bill and its implications, read Zach Williams’ brief in Bloomberg Law. To learn about the criticisms towards Virginia’s and Texas’ bill, read Adam Thierer’s analysis.

Links to Additional News

Industry:

• OpenAI unveils new image generator for ChatGPT (The New York Times).

• A federal judge in California denied Elon Musk’s attempt to block OpenAI from becoming a for-profit entity (CNBC).

• Meta faces legal challenge by French publishers for copyright infringement over AI training (Bloomberg).

• Anthropic wins early round in music publishers’ AI copyright case (Reuters).

• Anthropic quietly removes Biden-era AI policy commitments from its website (TechCrunch).

• Google reports scale of complaints about AI deepfake terrorism content to Australian regulator (Reuters).

• People are using Google’s new AI model to remove watermarks from images (TechCrunch).

• Google introduced Gemini 2.5 Pro Experimental as their most intelligent AI model (Google).

• Baidu releases reasoning AI model to take on DeepSeek (Bloomberg).

• DeepSeek releases V3 model update with significant improvements (Forbes).

• Claude can now search the web in feature preview for all paid users in the U.S. (Anthropic).

• NewsGuard launches new service to protect LLMs from foreign influence operations (NewsGuard).

• Amazon to Reportedly Process Alexa Voice Recordings in the Cloud, Ending Local Processing (Gadgets360).

• Italian newspaper says it has published world’s first AI-generated edition (The Guardian).

Government:

• China to crack down on stock market fake news as AI spurs misinformation, says state media (Reuters)

• UK lawmaker introduces AI Regulation bill in the House of Lords (PYMNTS)

• Spain to impose massive fines for not labelling AI-generated content (Reuters).

• New Texas bill could make incredibly popular anime & video games illegal (Dexerto)

• California’s AI experts release draft report of recommendations for AI guardrails (CalMatters).

• A well-funded Moscow-based global “news” network has infected Western artificial intelligence tools worldwide with Russian propaganda (NewsGuard)

• Leaked data exposes a Chinese AI censorship machine (TechCrunch)

• Hundreds of actors and Hollywood insiders sign open letter urging government not to loosen copyright laws for AI (CBS News)

The Future of Free Speech in Action

Founder and Executive Director Jacob Mchangama spoke at the Designing Trustworthy AI in a Polarized World conference hosted by the Hoover Institution and the Stanford Graduate School of Business, discussing free speech and political slant in AI.

In March, The Future of Free Speech also published The Future of Free Speech Index 2025, which found that global support for free-speech principles has been declining in recent years, even in democracies. The accompanying report shows that people are much less permissive with AI-generated content than with human-generated content.

We submitted comments to the White House’s Request for Information on implementing the AI Action Plan. We also provided feedback on the third draft of the EU’s General-Purpose AI Code of Practice.

Jordi Calvet-Bademunt is a Senior Research Fellow at The Future of Free Speech and a Visiting Scholar at Vanderbilt University. Isabelle Anzabi is a research associate at The Future of Free Speech, where she analyzes the intersections between AI policy and freedom of expression.

Share

TL;DR

• The Paris AI Action Summit’s Statement on Inclusive and Sustainable AI for People and Planet had broad international support with the notable exceptions of the U.S. and the UK. U.S. Vice President JD Vance’s speech warned against excessive regulation of AI and claimed that the administration is committed to “free speech.”

• The EU AI Act’s provisions on “unacceptable risk” in AI systems and AI literacy requirements have taken effect, with the Commission issuing non-binding guidelines on how to define “AI system” and on “unacceptable risk” prohibitions. The European Commission has withdrawn the AI Liability Directive due to a lack of foreseeable agreement.

• The U.S. Senate passed the TAKE IT DOWN Act, a bill that would criminalize non-consensual intimate imagery (NCII), including deepfakes, and mandate that social media remove such content within 48 hours. Free speech advocates have raised alarms regarding the removal obligation.

• The UK will introduce new AI-related sexual abuse offences, criminalizing AI-generated child sexual abuse material (CSAM) and AI “paedophile manuals,” alongside measures targeting CSAM websites and granting an enforcement authority the power to compel individuals to unlock their devices based on “reasonable suspicion.”

• South Korea removed DeepSeek, a Chinese AI chatbot, from app stores in the country due to data protection concerns.

Main Stories

The Paris AI Action Summit Redefines Global AI Agenda

France hosted the AI Action Summit, focusing on multilateral cooperation in AI innovation, investment, cultural impact, environmental sustainability, and inclusivity. It was the third global AI summit after ones held in the UK in 2023 and South Korea in 2024. The event brought together over 1,000 participants and several dozen heads of state.

A key outcome was the “Statement on Inclusive and Sustainable Artificial Intelligence for People and the Planet,” signed by more than 60 countries and regional blocs—notably excluding the U.S. and the UK. Meanwhile, U.S. Vice President JD Vance cautioned Action Summit attendees against “excessive regulation” of AI. He reaffirmed that the “Trump Administration will ensure that AI systems developed in America are free from ideological bias and never restrict [U.S.] citizens’ right to free speech.”

Jovan Kurbalija, from Diplo, explains that the Paris Action Summit shifted the global conversation on AI, “moving away from speculative long-term risks and toward immediate, tangible issues like innovation, jobs, and public good.” Unlike previous AI forums, this summit steered away from a safety-focused agenda, focusing on short-term, existing risks. For a deeper analysis of the impact of the Paris AI Action Summit, read Diplo’s brief.

In a separate event, the Munich Security Conference, Vice President Vance also claimed that Europe has gone too far in restricting speech. The Future of Free Speech Executive Director Jacob Mchangama recently argued that while Europe has a free speech problem, so does the Trump administration.

European Commission Withdraws the AI Liability Directive

The European Commission abandoned the EU AI Liability Directive proposal, arguing that there was “[n]o foreseeable agreement.” The Commission will now assess “whether another proposal should be tabled or another type of approach should be chosen.” The AI Liability Directive was initially proposed to “lay down uniform rules for certain aspects of non-contractual civil liability for damage caused with the involvement of AI systems.”

In late January, an industry coalition called for its withdrawal “to avoid further legal uncertainty and support Europe’s competitiveness.” The Commission’s decision follows criticism of the EU’s pro-regulatory approach to AI and aligns with the global shift toward pro-innovation policies, as seen at the Paris AI Action Summit. However, the European Parliament’s Internal Market and Consumer Protection Committee has voted to keep working on the proposal.

To learn more about the Commission’s decision to withdraw the AI Liability Directive, read IAPP’s brief.

EU AI Act’s Rules on Prohibited AI Practices Have Taken Effect

The EU AI Act’s provisions on “unacceptable risk” in AI systems became applicable on February 2. These provisions ban AI systems and practices “deemed unacceptable due to their potential risks to European values and fundamental rights.” This includes, for instance, the use of AI systems that deploy subliminal or manipulative techniques, predict the risk of criminal behavior of individuals, or infer emotions in the workplace or education. The European Commission also released non-binding guidelines to help stakeholders comply with the AI Act’s provisions on “unacceptable risk.”

The provisions on AI literacy requirements also took effect. They require providers and deployers of AI systems to ensure “to their best extent” that staff dealing with AI systems for business have sufficient skills to deploy AI systems and awareness about the risks and opportunities these systems present. The AI Office is expected to issue formal guidelines soon, but it has already published a Living Repository to Foster Learning and Exchange on AI Literacy.

Additionally, the Commission issued guidelines on the definition of AI systems, emphasizing that these guidelines “are designed to evolve over time and will be updated as necessary.” Like the compliance guidelines, these non-binding drafts assist developers in determining whether a software system qualifies as an AI system under the AI Act.

To learn more about the AI Act’s requirements coming into force, read Mayer Brown’s brief.

The TAKE IT DOWN Act, An Anti-Deepfake Law, Passed in The U.S. Senate

The U.S. Senate passed the TAKE IT DOWN Act (S. 146). This bill would make it a criminal offense to publish NCII, including AI-generated NCII (“deepfake revenge pornography”), and mandates that social media platforms and similar websites establish procedures to remove such content within 48 hours of receiving a victim’s notice. Advocates for abuse victims praised the move, but digital rights groups warn that the mandated 48-hour notice-and-takedown system is overbroad and threatens lawful speech. They argue the bill “is likely unconstitutional and will undoubtedly have a censorious impact on users’ free expression” since attempts to comply could result in the removal of satire, journalism, and other protected content. The bill is currently awaiting action in the House.

To learn more about the bill and its implications, read Kaylee Williams’ analysis in Tech Policy Press.

UK Targets AI-Generated Child Sex Abuse Images with New Laws

The United Kingdom will be the first country to create “new AI sexual abuse offences to protect children from predators generating AI images,” according to the Home Office. The new laws would make it illegal to possess, create, or distribute AI tools designed to generate child sexual abuse material and for anyone to possess AI “paedophile manuals” that teach people how to use AI to sexually abuse children. Additionally, the Home Office will introduce two other related offenses.

Additional measures criminalize predators who run websites designed for CSAM or grooming and allow Border Force to compel an individual to unlock their phone based on reasonable suspicion that they pose a sexual risk to children. These measures will be introduced as a part of the Crime and Policing Bill before Parliament in the spring. While these rules are undoubtedly well-intentioned and concern a very serious issue, they must be scrutinized to ensure that they respect the fundamental right to freedom of expression and are not excessively broad and vague.

To learn more about the UK’s new AI sexual abuse offenses, read the BBC’s brief.

South Korea Removes DeepSeek from App Stores in The Country

DeepSeek, a Chinese AI chatbot that quickly gained global popularity, has been removed from South Korean app stores by the country’s Personal Information Protection Commission (PIPC) over data protection concerns. The commission cited a lack of transparency around third‐party data transfers and excessive personal information collection, leading to a temporary ban on new downloads until the app meets stricter privacy standards. Italy also blocked DeepSeek for privacy reasons in January, and authorities in Australia, Taiwan, and several U.S. states have banned DeepSeek from government devices over data privacy and censorship concerns.

To learn more about the restrictions and security concerns, read the brief by Charles Mok, from the Global Digital Policy Incubator.

Links to Additional News

Industry:

• Apple teams up with Alibaba to bring AI features for iPhones in China (Reuters).

• Google and Meta are leading the charge against a “code of practice” governing tools like Gemini, ChatGPT and Llama (Politico).

• Google erases promise not to use AI technology for weapons or surveillance (CNN).

• Grok 3, X’s AI model, appears to have briefly censored unflattering mentions of Trump and Musk (TechCrunch).

• Meta says it may stop the development of frontier AI systems it deems too risky (TechCrunch).

• OpenAI tries to ‘uncensor’ ChatGPT (TechCrunch).

• OpenAI is facing a copyright lawsuit in India from prominent players in the music industry (The National Law Review).

• Scale AI and the U.S. AI Safety Institute are partnering to provide third-party evaluation in developing AI testing criteria and assessing AI models (Scale AI).

• Guardian Media Group announces strategic partnership with OpenAI that will provide access to Guardian’s editorial content through ChatGPT (The Guardian).

• Thomson Reuters Wins First Major AI Copyright Case in the US (Wired).

Government:

• The White House seeks public input on the Development of an Artificial Intelligence Action Plan with comments due March 15 (The Federal Register).

• The U.S. AI Safety Institute could face big cuts, with hundreds of staffers potentially laid off (TechCrunch).

• In the U.S., two recent federal court rulings on AI copyright infringement and fair use trended in favor of copyright holders and against AI companies (Transparency Coalition).

• The European Parliament Research Service published a brief on “Algorithmic discrimination under the AI Act and the GDPR” (European Parliament).

• EU launches InvestAI initiative to mobilize €200 billion (European Commission).

• France announced €109 billion investment into AI projects and data centers over the next several years (Le Monde).

• UK’s AI Safety Institute becomes the AI Security Institute, focusing on AI cybersecurity risks and not “bias or freedom of speech” (UK Government).

• UAE plans to launch new DeepSeek-inspired AI models, according to a senior official (AFP).

• Canada and Japan sign the Council of Europe’s first-ever global treaty on AI (Council of Europe).

The Future of Free Speech in Action

The Future of Free Speech hosted a RightsCon 2025 panel on “T&S in Gen AI: Agreeing on Principles for Freedom and Safety,” featuring panelists Sarah Shirazyan (Meta, Stanford), David Evan Harris (UC Berkeley), and Sayash Kapoor (Princeton), with Jordi Calvet-Bademunt (The Future of Free Speech) moderating.

Senior Research Fellow Jordi Calvet-Bademunt also spoke at the 2025 Tennessee Campus Civic Summit at the University of Tennessee Chattanooga, discussing the implications of AI for democracy.

Jordi Calvet-Bademunt is a Senior Research Fellow at The Future of Free Speech and a Visiting Scholar at Vanderbilt University. Isabelle Anzabi is a research associate at The Future of Free Speech, where she analyzes the intersections between AI policy and freedom of expression.

Share

The Future of Free Speech is excited to announce our new monthly newsletter, “.exe-pression.” Each month, Senior Research Fellow Jordi Calvet-Bademunt and Research Associate for AI Policy Isabelle Anzabi will explore the latest developments in global AI policy affecting free speech.

TL;DR

• In the United States, the Trump Administration reshaped the U.S. approach to AI policy by rescinding Biden’s Executive Order on AI and issuing a new Executive Order. In the meantime, Texas is considering risk-based legislation to govern AI.

• The European Union already has risk-based legislation, and the European Commission recently released the second draft of the General-Purpose AI Code of Practice. The first measures of the AI Act entered into effect in early February.

• South Korea passed a comprehensive AI regulatory framework. Brazil may be on track to follow suit, as its Senate has passed a similar bill.

• The Chinese company DeepSeek sent shockwaves throughout the AI world by launching a model that performed on par with flagship U.S. models while raising concerns about censorship.

Main Stories

The U.S. President Revokes AI Executive Order Reshaping AI Policy Approach

The U.S. is shifting its approach to AI Governance. Right after President Trump’s inauguration, he rescinded the “Executive Order on Safe, Secure, and Trustworthy Artificial Intelligence” issued by former President Biden in 2023. Just a few days later, President Trump issued a new Executive Order, "Removing Barriers to American Leadership in Artificial Intelligence,” which revokes existing, unidentified policies that act as barriers to American AI innovation.

The Order highlights the objective of developing systems “free from ideological bias or engineered social agendas.” What this means in practice and its implications for free speech remains to be seen.

The Executive Order requires a new AI action plan that sustains and enhances “America’s global AI dominance in order to promote human flourishing, economic competitiveness, and national security.” It also requires the removal of actions “inconsistent with, or present obstacles to” this objective.

European Commission Released the Second Draft of the General-Purpose AI Code of Practice

The EU promulgated its AI framework, the AI Act, in June 2024. In December 2024, the European Commission released the second draft of the General-Purpose AI Code of Practice. When approved, this Code will guide general-purpose AI providers in implementing the provisions of the AI Act on systemic risk and other requirements until harmonized standards are approved in a few years.

Over 1,000 people are developing the Code of Practice for Generative AI, a process led by five chairs supervising four working groups on issues like “risk identification and assessment” and “technical risk mitigation.” The final draft of the Code is expected to be presented in April 2025 and subsequently considered for approval. So far, the draft code of conduct has raised concerns about how some of its provisions – like one aimed at mitigating the risk of harmful manipulation – could impact free speech.

To understand why free-speech advocates should press for clear protections for freedom of expression in the AI Act implementation, you can read Jordi’s article: Safeguarding Freedom of Expression in the AI Era. To learn more about the Draft of the General-Purpose AI Code of Practice, you can read the European Commission’s official press release.

South Korea Passed the AI Basic Act

South Korea’s National Assembly recently passed the “Act on the Development of Artificial Intelligence and the Establishment of Trust” (AI Basic Act), which will enter into force in January 2026. Following the European Union, the AI Basic Act establishes the world's second comprehensive AI regulatory framework.

The law's stated purpose is to protect rights and dignity, enhance quality of life, and boost national competitiveness by establishing principles for AI development and trust. Notably, the AI Basic Act imposes transparency, safety, and reliability measures for “high-impact” AI. High-impact AI includes systems operating in 11 designated areas that could substantially affect or risk human life, physical safety, and fundamental rights.

The Act also mandates companies to label generative AI outputs and notify users when interacting with AI systems. Much like with the EU AI Act, free speech advocates should ensure that Korea’s AI Basic Act respects free speech and protects users’ right to access information.

To learn more about Korea’s new AI law, read IAPP’s brief: Analyzing South Korea's Framework Act on the Development of AI.

The Brazilian Senate Passed a Comprehensive AI Bill

Brazil could be the following country to establish an AI framework. In December 2024, the Brazilian Federal Senate passed Bill No. 2338/2023, which also aims to regulate the development and use of artificial intelligence, establishing a risk-based approach. The bill prohibits “excessive risk” AI systems and obligates high-risk systems to conduct algorithmic risk assessments, human supervision, and additional transparency measures.

The bill provides penalties for violations and creates the central overseeing authority, the National System for Artificial Intelligence and Governance. It also includes watermarking obligations and provisions aimed at protecting the fundamental rights of individuals and groups affected by AI systems. The Chamber of Deputies is now considering the bill.

To learn more about Brazil’s Bill, read Mattos Filho’s brief, Regulatory framework for artificial intelligence passes in Brazil’s Senate.

Texas Responsible AI Governance Act (TRAIGA) introduced in Texas Legislature

Texas is also considering a risk-based approach to govern AI. In December 2024, the Texas Responsible AI Governance Act was introduced into the Texas Legislature. It seeks to regulate the development and deployment of AI systems in the state. The bill would ban AI systems that pose unacceptable risk and establish obligations for high-risk AI systems.

To learn more about Texas’ Bill, read Orrick’s brief, The Texas Responsible AI Governance Act: 5 Things to Know.

DeepSeek’s AI Model Rivals Western Flagship Models

DeepSeek, a Chinese AI startup, launched an open-source AI model that rivals industry-leading models developed in the West, including OpenAI’s. The model quickly became one of the most downloaded apps across several markets, with millions of downloads.

DeepSeek claims it cost only $6 million to train the model, though this figure has been disputed. Despite DeepSeek’s promising performance, concerns about censorship, data privacy, and national security have been raised. Chinese AI companies must comply with strict government oversight and regulations on “sensitive topics,” which raises questions about the model’s ability to provide unbiased responses.

To learn more about DeepSeek’s new model, read The New York Times’ brief: What to Know About DeepSeek and How It Is Upending AI. The Guardian analyzed how the model presented propaganda and censored sensitive topics related to China: We tried out DeepSeek. It worked well, until we asked it about Tiananmen Square and Taiwan.

Links to Additional News

Industry:

• OpenAI plans to transition from a non-profit to for-profit structure to advance its mission

• Stargate Project intends to invest $500 billion to build new U.S. AI infrastructure

• Google pushes global agenda to educate workers and lawmakers on AI

Government:

• The U.S. House of Representatives Bipartisan AI Task Force releases a report offering key principles, findings, and proposals on AI

• New York Governor signs the legislative oversight of automated decision-making in government act (LOADinG Act)

• The Pope issues the Guidelines on AI, establishing ethical principles for the development and use of AI in the Vatican City State

• Biden-Harris Administration expanded export controls for advanced computing chips and certain closed AI model weights

Research:

• MLCommons’ AILuminate v1.0 benchmark release provides safety testing for general-purpose chat systems across twelve hazard categories

• A new paper from Anthropic provides the first empirical example of an LLM engaging in alignment faking without having been trained or instructed to do so

• Stanford Law Review article provides speech certainty as an approach to conceptualizing machine learning algorithms under existing First Amendment jurisprudence.

The Future of Free Speech in Action

The Future of Free Speech will be hosting a RightsCon 2025 panel on “T&S in Gen AI: Agreeing on Principles for Freedom and Safety.” The discussion will be held online on February 26 and will feature panelists Sarah Shirazyan (Meta, Stanford), David Evan Harris (UC Berkeley), and Sayash Kapoor (Princeton), with Jordi Calvet-Bademunt (The Future of Free Speech) moderating. You can register for the event here.

Jordi Calvet-Bademunt will be speaking at the 2025 Tennessee Campus Civic Summit on February 28 at the University of Tennessee Chattanooga, discussing the implications of AI for democracy. Attendance is in person, and you can register here.

Jordi Calvet-Bademunt is a Senior Research Fellow at The Future of Free Speech and a Visiting Scholar at Vanderbilt University. Isabelle Anzabi is a research associate at The Future of Free Speech, where she analyzes the intersections between AI policy and freedom of expression.

Share The Bedrock Principle